An Introduction To Machine Learning 2nd Ed

Miroslav
Kubat
An Introduction
to Machine
Learning
Second Edition
An Introduction to Machine Learning
Miroslav Kubat
An Introduction to Machine
Learning
Second Edition
123
Miroslav Kubat
Department of Electrical and Computer Engineering
University of Miami
Coral Gables, FL, USA
ISBN 978-3-319-63912-3 ISBN 978-3-319-63913-0 (eBook)

DOI 10.1007/978-3-319-63913-0
Library of Congress Control Number: 2017949183
© Springer International Publishing AG 2015, 2017

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
Printed on acid-free paper
This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
To my wife, Verunka.
Contents
1 A Simple Machine-Learning Task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Training Sets and Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Minor Digression: Hill-Climbing Search . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Hill Climbing in Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 The Induced Classifier’s Performance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.5 Some Difficulties with Available Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.6 Summary and Historical Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.7 Solidify Your Knowledge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2 Probabilities: Bayesian Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1 The Single-Attribute Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2 Vectors of Discrete Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3 Probabilities of Rare Events: Exploiting the Expert’s Intuition . . . . 26
2.4 How to Handle Continuous Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.5 Gaussian “Bell” Function: A Standard pdf . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.6 Approximating PDFs with Sets of Gaussians . . . . . . . . . . . . . . . . . . . . . . . 34
3 Similarities: Nearest-Neighbor Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.1 The k-Nearest-Neighbor Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.2 Measuring Similarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.3 Irrelevant Attributes and Scaling Problems . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.4 Performance Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.5 Weighted Nearest Neighbors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.6 Removing Dangerous Examples. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.7 Removing Redundant Examples. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
vii
viii Contents
4 Inter-Class Boundaries: Linear and Polynomial Classifiers . . . . . . . . . . . 65

4.1 The Essence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2 The Additive Rule: Perceptron Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.3 The Multiplicative Rule: WINNOW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.4 Domains with More Than Two Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.5 Polynomial Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.6 Specific Aspects of Polynomial Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.7 Numerical Domains and Support Vector Machines . . . . . . . . . . . . . . . . 84
5 Artificial Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.1 Multilayer Perceptrons as Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.2 Neural Network’s Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.3 Backpropagation of Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.4 Special Aspects of Multilayer Perceptrons. . . . . . . . . . . . . . . . . . . . . . . . . . 100
5.5 Architectural Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.6 Radial-Basis Function Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6 Decision Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.1 Decision Trees as Classifiers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.2 Induction of Decision Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.3 How Much Information Does an Attribute Convey? . . . . . . . . . . . . . . . 119
6.4 Binary Split of a Numeric Attribute . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.5 Pruning. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
6.6 Converting the Decision Tree into Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
7 Computational Learning Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
7.1 PAC Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
7.2 Examples of PAC Learnability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
7.3 Some Practical and Theoretical Consequences . . . . . . . . . . . . . . . . . . . . . 143
7.4 VC-Dimension and Learnability. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
7.6 Exercises and Thought Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
8 A Few Instructive Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
8.1 Character Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
8.2 Oil-Spill Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
8.3 Sleep Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
8.4 Brain–Computer Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
8.5 Medical Diagnosis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Contents ix
8.6 Text Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

8.8 Exercises and Thought Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
9 Induction of Voting Assemblies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
9.1 Bagging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
9.2 Schapire’s Boosting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
9.3 Adaboost: Practical Version of Boosting . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
9.4 Variations on the Boosting Theme. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
9.5 Cost-Saving Benefits of the Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
10 Some Practical Aspects to Know About. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
10.1 A Learner’s Bias. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
10.2 Imbalanced Training Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
10.3 Context-Dependent Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
10.4 Unknown Attribute Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
10.5 Attribute Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
10.6 Miscellaneous . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
11 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
11.1 Basic Performance Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
11.2 Precision and Recall. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
11.3 Other Ways to Measure Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
11.4 Learning Curves and Computational Costs . . . . . . . . . . . . . . . . . . . . . . . . . 222
11.5 Methodologies of Experimental Evaluation . . . . . . . . . . . . . . . . . . . . . . . . 224
12 Statistical Significance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
12.1 Sampling a Population . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
12.2 Benefiting from the Normal Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
12.3 Confidence Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
12.4 Statistical Evaluation of a Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
12.5 Another Kind of Statistical Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
12.6 Comparing Machine-Learning Techniques . . . . . . . . . . . . . . . . . . . . . . . . . 245
13 Induction in Multi-Label Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
13.1 Classical Machine Learning in Multi-Label Domains . . . . . . . . . . . . . . 251
13.2 Treating Each Class Separately: Binary Relevance . . . . . . . . . . . . . . . . . 254
13.3 Classifier Chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
x Contents
13.4 Another Possibility: Stacking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258

13.5 A Note on Hierarchically Ordered Classes. . . . . . . . . . . . . . . . . . . . . . . . . . 260
13.6 Aggregating the Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
13.7 Criteria for Performance Evaluation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
14 Unsupervised Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
14.1 Cluster Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
14.2 A Simple Algorithm: k-Means. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
14.3 More Advanced Versions of k-Means . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
14.4 Hierarchical Aggregation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
14.5 Self-Organizing Feature Maps: Introduction. . . . . . . . . . . . . . . . . . . . . . . . 286
14.6 Some Important Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
14.7 Why Feature Maps? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
15 Classifiers in the Form of Rulesets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
15.1 A Class Described By Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
15.2 Inducing Rulesets by Sequential Covering . . . . . . . . . . . . . . . . . . . . . . . . . . 300
15.3 Predicates and Recursion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
15.4 More Advanced Search Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
16 The Genetic Algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
16.1 The Baseline Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
16.2 Implementing the Individual Modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
16.3 Why It Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
16.4 The Danger of Premature Degeneration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
16.5 Other Genetic Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
16.6 Some Advanced Versions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
16.7 Selections in k-NN Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
17 Reinforcement Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
17.1 How to Choose the Most Rewarding Action . . . . . . . . . . . . . . . . . . . . . . . . 331
17.2 States and Actions in a Game . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
17.3 The SARSA Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
Introduction
Machine learning has come of age. And just in case you might think this is a mere
platitude, let me clarify.
The dream that machines would one day be able to learn is as old as computers
themselves, perhaps older still. For a long time, however, it remained just that: a
dream. True, Rosenblatt’s perceptron did trigger a wave of activity, but in retrospect,
the excitement has to be deemed short-lived. As for the attempts that followed, these
fared even worse; barely noticed, often ignored, they never made a breakthrough—
no software companies, no major follow-up research, and not much support from
funding agencies. Machine learning remained an underdog, condemned to live in
the shadow of more successful disciplines. The grand ambition lay dormant.
And then it all changed.
A group of visionaries pointed out a weak spot in the knowledge-based systems
that were all the rage in the 1970s’ artificial intelligence: where was the “know-
ledge” to come from? The prevailing wisdom of the day insisted that it should
take the form of if-then rules put together by the joint effort of engineers and field
experts. Practical experience, though, was unconvincing. Experts found it difficult
to communicate what they knew to engineers. Engineers, in turn, were at a loss as
to what questions to ask and what to make of the answers. A few widely publicized
success stories notwithstanding, most attempts to create a knowledge base of, say,
tens of thousands of such rules proved frustrating.
The proposition made by the visionaries was both simple and audacious. If it is
so hard to tell a machine exactly how to go about a certain problem, why not provide
the instruction indirectly, conveying the necessary skills by way of examples from
which the computer will—yes—learn!
Of course, this only makes sense if we can rely on the existence of algorithms to
do the learning. This was the main difficulty. As it turned out, neither Rosenblatt’s
perceptron nor the techniques developed after it were very useful. But the absence
of the requisite machine-learning techniques was not an obstacle; rather, it was a
challenge that inspired quite a few brilliant minds. The idea of endowing computers
with learning skills opened new horizons and created a large amount of excitement.
The world was beginning to take notice.
xi
xii Introduction
The bombshell exploded in 1983. Machine Learning: The AI Approach1 was

a thick volume of research papers which proposed the most diverse ways of
addressing the great mystery. Under their influence, a new scientific discipline
was born—virtually overnight. Three years later, a follow-up book appeared and
then another. A soon-to-become-prestigious scientific journal was founded. Annual
conferences of great repute were launched. And dozens, perhaps hundreds, of
doctoral dissertations, were submitted and successfully defended.
In this early stage, the question was not only how to learn but also what to learn
and why. In retrospect, those were wonderful times, so creative that they deserve to
be remembered with nostalgia. It is only to be regretted that so many great thoughts
later came to be abandoned. Practical needs of realistic applications got the upper
hand, pointing to the most promising avenues for further efforts. After a period of
enchantment, concrete research strands crystallized: induction of the if-then rules for
knowledge-based systems; induction of classifiers, programs capable of improving
their skills based on experience; automatic fine-tuning of Prolog programs; and
some others. So many were the directions that some leading personalities felt
it necessary to try to steer further development by writing monographs, some
successful, others less so.
An important watershed was Tom Mitchell’s legendary textbook.2 This summa-
rized the state of the art of the field in a format appropriate for doctoral students
and scientists alike. One by one, universities started offering graduate courses that
were usually built around this book. Meanwhile, the research methodology became
more systematic, too. A rich repository of machine-leaning test beds was created,
making it possible to compare the performance or learning algorithms. Statistical
methods of evaluation became widespread. Public domain versions of most popular
programs were made available. The number of scientists dealing with this discipline
grew to thousands, perhaps even more.
Now, we have reached the stage where a great many universities are offering
machine learning as an undergraduate class. This is quite a new situation. As a
rule, these classes call for a different kind of textbook. Apart from mastering the
baseline techniques, future engineers need to develop a good grasp of the strengths
and weaknesses of alternative approaches; they should be aware of the peculiarities
and idiosyncrasies of different paradigms. Above all, they must understand the
circumstances under which some techniques succeed and others fail. Only then will
they be able to make the right choices when addressing concrete applications. A
textbook that is to provide all of the above should contain less mathematics, but a
lot of practical advice.
These then are the considerations that have dictated the size, structure, and style
of a teaching text meant to provide the material for a one-semester introductory
course.
1
Edited by R. Michalski, J. Carbonell, and T. Mitchell.
2
T. Mitchell, Machine Learning, McGraw-Hill (1997).
Introduction xiii
The first problem is the choice of material. At a time when high-tech companies
are establishing machine-learning groups, universities have to provide the students
with such knowledge, skills, and understanding that are relevant to the current needs
of the industry. For this reason, preference has been given to Bayesian classifiers,
nearest-neighbor classifiers, linear and polynomial classifiers, decision trees, the
fundamentals of the neural networks, and the principle of the boosting algorithms.
Significant space has been devoted to certain typical aspects of concrete engineering
applications. When applied to really difficult tasks, the baseline techniques are
known to behave not exactly the same way they do in the toy domains employed
by the instructor. One has to know what to expect.
The book consists of 17 chapters, each covering one major topic. The chapters are
divided into sections, each devoted to one critical problem. The student is advised
to proceed to the next section only after having answered the set of 2–4 “control
questions” at the end of the previous section. These questions are here to help
the student decide whether he or she has mastered the given material. If not, it is
necessary to return to the previous text.
As they say, only practice makes perfect. This is why at the end of each chapter
are exercises to encourage the necessary practicing. Deeper insight into the diverse
aspects of the material will then be gained by going through the thought experiments
that follow. These are more difficult, but it is only through hard work that an
engineer develops the right kind of understanding. The acquired knowledge is then
further solidified by suggested computer projects. Programming is important, too.
Nowadays, everybody is used to downloading the requisite software from the web.
This shortcut, however, is not recommended to the student of this book. It is only
by being forced to flesh out all the details of a computer program that you learn to
appreciate all the subtle points of the machine-learning techniques presented here.
Chapter 1
A Simple Machine-Learning Task
You will find it difficult to describe your mother’s face accurately enough for your
friend to recognize her in a supermarket. But if you show him a few of her photos,
he will immediately spot the tell-tale traits he needs. As they say, a picture—an
example—is worth a thousand words.
This is what we want our technology to emulate. Unable to define certain objects
or concepts with adequate accuracy, we want to convey them to the machine by
way of examples. For this to work, however, the computer has to be able to convert
the examples into knowledge. Hence our interest in algorithms and techniques for
machine learning, the topic of this textbook.
The first chapter formulates the task as a search problem, introducing hill-
climbing search not only as our preliminary attempt to address the machine-learning
task, but also as a tool that will come handy in a few auxiliary problems to be
encountered in later chapters. Having thus established the foundation, we will
proceed to such issues as performance criteria, experimental methodology, and
certain aspects that make the learning process difficult—and interesting.
1.1 Training Sets and Classifiers
Let us introduce the problem, and certain fundamental concepts that will accompany
us throughout the rest of the book.
The Set of Pre-Classified Training Examples Figure 1.1 shows six pies that
Johnny likes, and six that he does not. These positive and negative examples of the
underlying concept constitute a training set from which the machine is to induce a
classifier—an algorithm capable of categorizing any future pie into one of the two
classes: positive and negative.
© Springer International Publishing AG 2017 1

M. Kubat, An Introduction to Machine Learning,
DOI 10.1007/978-3-319-63913-0_1
2 1 A Simple Machine-Learning Task
Johnny likes:
Johnny does NOT like:
Fig. 1.1 A simple machine-learning task: induce a classifier capable of labeling future pies as
positive and negative instances of “a pie that Johnny likes”
The number of classes can of course be greater. Thus a classifier that decides
whether a landscape snapshot was taken in spring, summer, fall, or
winter distinguishes four. Software that identifies characters scribbled on an
iPad needs at least 36 classes: 26 for letters and 10 for digits. And document-
categorization systems are capable of identifying hundreds, even thousands of
different topics. Our only motivation for choosing a two-class domain is its
simplicity.
1.1 Training Sets and Classifiers 3
Table 1.1 The twelve training examples expressed in a matrix form

Crust Filling
Example Shape Size Shade Size Shade Class
ex1 Circle Thick Gray Thick Dark pos
ex2 Circle Thick White Thick Dark pos
ex3 Triangle Thick Dark Thick Gray pos
ex4 Circle Thin White Thin Dark pos
ex5 Square Thick Dark Thin White pos
ex6 Circle Thick White Thin Dark pos
ex7 Circle Thick Gray Thick White neg
ex8 Square Thick White Thick Gray neg
ex9 Triangle Thin Gray Thin Dark neg
ex10 Circle Thick Dark Thick White neg
ex11 Square Thick White Thick Dark neg
ex12 Triangle Thick White Thick Gray neg
Attribute Vectors To be able to communicate the training examples to the

machine, we have to describe them in an appropriate way. The most common
mechanism relies on so-called attributes. In the “pies” domain, five may be
suggested: shape (circle, triangle, and square), crust-size (thin or thick),
crust-shade (white, gray, or dark), filling-size (thin or thick), and
filling-shade (white, gray, or dark). Table 1.1 specifies the values of these
attributes for the twelve examples in Fig. 1.1. For instance, the pie in the upper-
left corner of the picture (the table calls it ex1) is described by the following
conjunction:
(shape=circle) AND (crust-size=thick) AND (crust-shade=gray)
AND (filling-size=thick) AND (filling-shade=dark)
A Classifier to Be Induced The training set constitutes the input from which we
are to induce the classifier. But what classifier?
Suppose we want it in the form of a boolean function that is true for
positive examples and false for negative ones. Checking the expression
[(shape=circle) AND (filling-shade=dark)] against the training
set, we can see that its value is false for all negative examples: while it is possible
to find negative examples that are circular, none of these has a dark filling. As for
the positive examples, however, the expression is true for four of them and false for
the remaining two. This means that the classifier makes two errors, a transgression
we might refuse to tolerate, suspecting there is a better solution. Indeed, the reader
will easily verify that the following expression never goes wrong on the entire
training set:
[ (shape=circle) AND (filling-shade=dark) ] OR
[ NOT(shape=circle) AND (crust-shade=dark) ]
Problems with a Brute-Force Approach How does a machine find a classifier of

this kind? Brute force (something that computers are so good at) will not do here.
Just consider how many different examples can be distinguished by the given set
of attributes in the “pies” domain. For each of the three different shapes, there
are two alternative crust-sizes, the number of combinations being 3 2 D 6.
For each of these, the next attribute, crust-shade, can acquire three different
values, which brings the number of combinations to 3 2 3 D 18. Extending this
line of reasoning to all attributes, we realize that the size of the instance space is
3 2 3 2 3 D 108 different examples.
Each subset of these examples—and there are 2108 subsets!—may constitute the
list of positive examples of someone’s notion of a “good pie.” And each such subset
can be characterized by at least one boolean expression. Running each of these
classifiers through the training set is clearly out of the question.
Manual Approach and Search Uncertain about how to invent a classifier-
inducing algorithm, we may try to glean some inspiration from an attempt
to create a classifier “manually,” by the good old-fashioned pencil-and-paper
method. When doing so, we begin with some tentative initial version, say,
shape=circular. Having checked it against the training set, we find it to
be true for four positive examples, but also for two negative ones. Apparently,
the classifier needs to be “narrowed” (specialized) so as to exclude the two
negative examples. One way to go about the specialization is to add a conjunction,
such as when turning shape=circular into [(shape=circular) AND
(filling-shade=dark)]. This new expression, while false for all negative
examples, is still imperfect because it covers only four (ex1, ex2, ex4, and
ex6) of the six positive examples. The next step should therefore attempt some
generalization, perhaps by adding a disjunction: {[(shape=circular) AND
(filling-shade=dark)] OR (crust-size=thick)}. We continue in
this way until we find a 100% accurate classifier (if it exists).
The lesson from this little introspection is that the classifier can be created by
means of a sequence of specialization and generalization steps which gradually
modify a given version of the classifier until it satisfies certain predefined require-
ments. This is encouraging. Readers with background in Artificial Intelligence will
recognize this procedure as a search through the space of boolean expressions. And
Artificial Intelligence is known to have developed and explored quite a few of search
algorithms. It may be an idea to take a look at least at one of them.
What Have You Learned?
To make sure you understand the topic, try to answer the following questions. If
needed, return to the appropriate place in the text.
• What is the input and output of the learning problem we have just described?
• How do we describe the training examples? What is instance space? Can we
calculate its size?
1.2 Minor Digression: Hill-Climbing Search 5
• In the “pies” domain, find a boolean expression that correctly classifies all the
training examples from Table 1.1.
1.2 Minor Digression: Hill-Climbing Search
Let us now formalize what we mean by search, and then introduce one pop-
ular algorithm, the so-called hill climbing. Artificial Intelligence defines search
something like this: starting from an initial state, find a sequence of steps which,
proceeding through a set of interim search states, lead to a predefined final state.
The individual steps—transitions from one search state to another—are carried out
by search operators which, too, have been pre-specified by the programmer. The
order in which the search operators are applied follows a specific search strategy
(Fig. 1.2).
Hill Climbing: An Illustration One popular search strategy is hill climbing. Let
us illustrate its essence on a well-known brain-teaser, the sliding-tiles puzzle. The
board of a trivial version of this game consists of nine squares arranged in three
rows, eight covered by numbered tiles (integers from 1 to 8), the last left empty. We
convert one search state into another by sliding to the empty square a tile from one
of its neighbors. The goal is to achieve a pre-specified arrangement of the tiles.
Search Operators Search Strategy
Initial State Final State
Search Agent
Fig. 1.2 A search problem is characterized by an initial state, final state, search operators, and a
search strategy
The flowchart in Fig. 1.3 starts with a concrete initial state, in which we can
choose between two operators: “move tile-6 up” and “move tile-2 to the
left.” The choice is guided by an evaluation function that estimates for each state
its distance from the goal. A simple possibility is to count the squares that the tiles
have to traverse before reaching their final destinations. In the initial state, tiles 2, 4,
and 5 are already in the right locations; tile 3 has to be moved by four squares; and
each of the tiles 1, 6, 7, and 8 have to be moved by two squares. This sums up to
distance d D 4 C 4 2 D 12.
In Fig. 1.3, each of the two operators applicable to the initial state leads to a
state whose distance from the final state is d D 13. In the absence of any other
guidance, we choose randomly and go to the left, reaching the situation where the
empty square is in the middle of the top row. Here, three moves are possible. One of
them would only get us back to the initial state, and can thus be ignored; as for the
remaining two, one results in a state with d D 14, the other in a state with d D 12.
The latter being the lower value, this is where we go. The next step is trivial because
only one move gets us to a state that has not been visited before. After this, we again
face the choice between two alternatives . . . and this how the search continues until
it reaches the final state.
Alternative Termination Criteria and Evaluation Functions Other termination
criteria can be considered, too. The search can be instructed to stop when the
maximum allotted time has elapsed (we do not want the computer to run forever),
when the number of visited states has exceeded a certain limit, when something
sufficiently close to the final state has been found, when we have realized that
all states have already been visited, and so on, the concrete formulation reflecting
critical aspects of the given application, sometimes combining two or more criteria
in one.
By the way, the evaluation function employed in the sliding-tiles example was
fairly simple, barely accomplishing its mission: to let the user convey some notion
of his or her understanding of the problem, to provide a hint as to which move a
human solver might prefer. To succeed in a realistic application, we would have to
come up with a more sophisticated function. Quite often, many different alternatives
can be devised, each engendering a different sequence of steps. Some will be quick
in reaching the solution, others will follow a more circuitous path. The program’s
performance will then depend on the programmer’s ability to pick the right one.
The Algorithm of Hill Combing The algorithm is summarized by the pseudocode
in Table 1.2. Details will of course depend on each individual’s programming style,
but the code will almost always contain a few typical functions. One of them
compares two states and returns true if they are identical; this is how the program
ascertains that the final state has been reached. Another function takes a given search
state and applies to it all search operators, thus creating a complete set of “child
states.” To avoid infinite loops, a third function checks whether a state has already
been investigated. A fourth calculates for a given state its distance from the final
1.2 Minor Digression: Hill-Climbing Search 7
Hill Climbing Final State:
1 1 2 3
2 1 8 4
initial state
6 7 4 7 6 5
3 8 5
2 d = 12
2 1 6 2 1
6 7 4 7 4
3 8 5 3 8 5
d = 13 d = 13
3
2 7 1 2 1
6 4 6 7 4
3 8 5 3 8 5
d = 14
d = 12
4
2 1 4
6 7
3 8 5
d = 13
5
2 1 4 2 1 4
6 7 6 7 5
3 8 5 3 8
d = 14 d = 14
Fig. 1.3 Hill climbing. Circled integers indicate the order in which the search states are visited.
d is a state’s distance from the final state as calculated by the given evaluation function. Ties are
broken randomly
state, and a fifth sorts the “child” states according to the distances thus calculated
and places them at the front of the list L. And the last function checks if a termination
criterion has been satisfied.1
One last observation: at some of the states in Fig. 1.3, no “child” offers any
improvement over its “parent,” a lower d-value being achieved only after temporary
compromises. This is what a mountain climber may experience, too: sometimes,
he has to traverse a valley before being able to resume the ascent. The mountain-
climbing metaphor, by the way, is what gave this technique its name.
1
For simplicity, the pseudocode ignores termination criteria other than reaching, or failing to reach,
the final state.
Table 1.2 Hill-climbing search algorithm

1. Create two lists, L and Lseen . At the beginning, L contains only the initial state, and Lseen is
empty.
2. Let n be the first element of L. Compare this state with the final state. If they are identical,
stop with success.
3. Apply to n all available search operators, thus obtaining a set of new states. Discard those
states that already exist in Lseen . As for the rest, sort them by the evaluation function and
place them at the front of L.
4. Transfer n from L into the list, Lseen , of the states that have been investigated.
5. If L D ;, stop and report failure. Otherwise, go to 2.
• How does Artificial Intelligence define the search problem? What do we
understand under the terms, “search space” and “search operators”?
• What is the role of the evaluation function? How does it affect the hill-climbing
behavior?
1.3 Hill Climbing in Machine Learning
We are ready to explore the concrete ways of applying hill climbing to the needs of
machine learning.
Hill Climbing and Johnny’s Pies Let us begin with the problem of how to decide
which pies Johnny likes. The input consists of a set of training examples, each
described by the available attributes. The output—the final state—is a boolean
expression that is true for each positive example in the training set, and false for each
negative example. The expression involves attribute-value pairs, logical operators
(conjunction, disjunction, and negation), and such combination of parentheses as
may be needed. The evaluation function measures the given expression’s error rate
on the training set. For the initial state, any randomly generated expression can be
used. In Fig. 1.4, we chose (shape=circle), on the grounds that more than a
half of the training examples are circular.
As for the search operator, one possibility is to add a conjunction as
illustrated in the upper part of Fig. 1.4: for instance, the root’s leftmost child
is obtained by replacing (shape=circle) with [(shape=circle) AND
(filling-shade=dark)] (in the picture, logical AND is represented by the
symbol “^.”). Note how many different expressions this operator generates even
in our toy domain. To shape=circle, any other attribute-value pair can be
1.3 Hill Climbing in Machine Learning 9
1 shape = circle
e = 4/12
Addition of (AND)
2
shape = circle shape = circle shape = circle
fill shade = dark fill size = thick crust size = thick
e = 2/12 e = 6/12 e = 5/12
Addition of (OR)
3
( shape = circle (shape = circle ( shape = circle
fill shade = dark ) fill shade = dark ) fill shade = dark )
shape = triangle crust shade = dark fill size = thin
e = 3/12 e = 1/12 e = 2/12
Fig. 1.4 Hill-climbing search in the “pies” domain
“ANDed.” Since the remaining four attributes (apart from shape) acquire 2, 3, 2,
and 3 different values, respectively, the total number of terms that can be added to
(shape=circle) is 2 2 3 D 36.2
Alternatively, we may choose to add a disjunction, as illustrated (in the picture)
by the three expansions of the leftmost child. Other operators may “remove a
conjunct,” “remove a disjunct,” “add a negation,” “negate a term,” various ways
of manipulating parentheses, and so on. All in all, hundreds of search operators can
be applied to each state, and then again to the resulting states. This can be hard to
manage even in this very simple domain.
Numeric Attributes In the “pies” domain, each attribute acquires one out of a
few discrete values, but in realistic applications, some attributes will probably be
numeric. For instance, each pie has a price, an attribute whose values come from
a continuous domain. What will the search look like then?
To keep things simple, suppose there are only two attributes: weight and
price. This limitation makes it possible, in Fig. 1.5, to represent each training
example by a point in a plane. The reader can see that examples belonging to
the same class tend to occupy a specific region, and curves separating individual
regions can be defined—expressed mathematically as lines, circles, polynomials.
For instance, the right part of Fig. 1.5 shows three different circles, each of which
can act as a classifier: examples inside the circle are deemed positive; those outside,
negative. Again, some of these classifiers are better than others. How will hill
climbing go about finding the best ones? Here is one possibility.
2
Of the 36 new states thus created, Fig. 1.4 shows only three.
Price Price
($3, 1.2lb) ($3, 1.2lb)
Weight Weight
Fig. 1.5 On the left: a domain with continuous attributes; on the right: some “circular” classifiers
Hill Climbing in a Domain with Numeric Attributes

Initial State A circle is defined by its center and radius. We can identify the initial
center with a randomly selected positive example, making the initial radius so small
that the circle contains only this single example.
Search Operators Two search operators can be used: one increases the circle’s
radius, and the other shifts the center from one training example to another. In the
former, we also have to determine how much the radius should change. One idea is
to increase it only so much as to make the circle encompass one additional training
example. At the beginning, only one training example is inside. After the first step,
there will be two, then three, four, and so on.
Final State The circle may not be an ideal figure to represent the positive region. In
this event, a 100% accuracy may not be achievable, and we may prefer to define the
final state as, say, a “classifier that correctly classifies 95% of the training examples.”
Evaluation Function As before, we choose to minimize the error rate.
• What aspects of search must be specified before we can employ hill climbing in
machine learning?
• What search operators can be used in the “pies” domain and what in the “circles”
domain? How can we define the evaluation function, the initial state, and the final
state?
1.4 The Induced Classifier’s Performance 11
1.4 The Induced Classifier’s Performance
So far, we have measured the error rate by comparing the training examples’ known
classes with those recommended by the classifier. Practically speaking, though, our
goal is not to re-classify objects whose classes we already know; what we really
want is to label future examples, those of whose classes we are as yet ignorant.
The classifier’s anticipated performance on these is estimated experimentally. It is
important to know how.
Independent Testing Examples The simplest scenario will divide the available
pre-classified examples into two parts: the training set, from which the classifier is
induced, and the testing set, on which it is evaluated (Fig. 1.6). Thus in the “pies”
domain, with its 12 pre-classified examples, the induction may be carried out on
randomly selected eight, and the testing on the remaining four. If the classifier then
“guesses” correctly the class of three testing examples (while going wrong on one),
its performance is estimated as 75%.
Reasonable though this approach may appear, it suffers from a major drawback:
a random choice of eight training examples may not be sufficiently representative
of the underlying concept—and the same applies to the (even smaller) testing set.
If we induce the meaning of a mammal from a training set consisting of a whale,
a dolphin, and a platypus, the learner may be led to believe that mammals live in
the sea (whale, dolphin), and sometimes lay eggs (platypus), hardly an opinion a
biologist will embrace. And yet, another choice of trainingexamples may result in a
Fig. 1.6 Pre-classified

available examples
examples are divided into the
training and testing sets
training testing
set set
classifier satisfying the highest standards. The point is, a different training/testing set
division gives rise to a different classifier—and also to a different estimate of future
performance. This is particularly serious if the number of pre-classified examples is
small.
Suppose we want to compare two machine learning algorithms in terms of the
quality of the products they induce. The problem of non-representative training
sets can be mitigated by so-called random subsampling.3 The idea is to repeat the
random division into the training and testing sets several times, always inducing a
classifier from the i-th training set, and then measuring the error rate, Ei , on the i-th
testing set. The algorithm that delivers classifiers with the lower average value of
Ei ’s is deemed better—as far as classification performance is concerned.
3
Later, we will describe some other methodologies.
The Need for Explanations In some applications, establishing the class of each
example is not enough. Just as desirable is to know the reasons behind the
classification. Thus a patient is unlikely to give consent to amputation if the only
argument in support of surgery is, “this is what our computer says.” But how to find
a better explanation?
In the “pies” domain, a lot can be gleaned from the boolean expression itself. For
instance, we may notice that a pie was labeled as negative whenever its shape was
square, and its filling white. Combining this observation with alternative sources
of knowledge may offer useful insights: the dark shade of the filling may indicate
poppy, an ingredient Johnny is known to love; or the crust of circular pies turns out
to be more crispy than that of square ones; and so on. The knowledge obtained in
this manner can be more desirable than the classification itself.
By contrast, the classifier in the “circles” domain is a mathematical expression
that acts as a “black box” which accepts an example’s description and returns the
class label without telling us anything else. This is not necessarily a shortcoming.
In some applications, an explanation is nothing more than a welcome bonus; in
others, it is superfluous. Consider a classifier that accepts a digital image of a hand-
written character and returns the letter it represents. The user who expects several
pages of text to be converted into a Word document will hardly insist on a detailed
explanation for each single character.
Existence of Alternative Solutions By the way, we should notice that many
apparently perfect classifiers can be induced from the given data. In the “pies”
domain, the training set contained 12 examples, and the classes of the remaining 96
examples were unknown. Using some simple combinatorics, we realize that there
are 296 classifiers that label correctly all training examples but differ in the way they
label the unknown 96. One induced classifier may label correctly every single future
example—and another will misclassify them all.
• How can we estimate the error rate on examples that have not been seen during
learning?
• Why is error rate usually higher on the testing set than on the training set?
• Give an example of a domain where the classifier also has to explain its action,
and an example of a domain where this is unnecessary.
• What do we mean by saying that, “there is a combinatorial number of classifiers
that correctly classify all training examples”?
1.5 Some Difficulties with Available Data 13
1.5 Some Difficulties with Available Data
In some applications, the training set is created manually: an expert prepares the
examples, tags them with class labels, chooses the attributes, and specifies the value
of each attribute in each example. In other domains, the process is computerized. For
instance, a company may want to be able to anticipate an employee’s intention to
leave. Their database contains, for each person, the address, gender, marital status,
function, salary raises, promotions—as well as the information about whether the
person is still with the company or, if not, the day they left. From this, a program
can obtain the attribute vectors, labeled as positive if the given person left within a
year since the last update of the database record.
Sometimes, the attribute vectors are automatically extracted from a database, and
labeled by an expert. Alternatively, some examples can be obtained from a database,
and others added manually. Often, two or more databases are combined. The number
of such variations is virtually unlimited.
But whatever the source of the examples, they are likely to suffer from imperfec-
tions whose essence and consequences the engineer has to understand.
Irrelevant Attributes To begin with, some attributes are important, while others
are not. While Johnny may be truly fond of poppy filling, his preference for a pie
will hardly be driven by the cook’s shoe size. This is something to be concerned
about: irrelevant attributes add to computational costs; they can even mislead the
learner. Can they be avoided?
Usually not. True, in manually created domains, the expert is supposed to know
which attributes really matter, but even here, things are not so simple. Thus the
author of the “pies” domain might have done her best to choose those attributes
she believed to matter. But unsure about the real reasons behind Johnny’s tastes,
she may have included attributes whose necessity she suspected—but could not
guarantee. Even more often the problems with relevance occur when the examples
are extracted from a database. Databases are developed primarily with the intention
to provide access to lots of information—of which usually only a tiny part pertains
to the learning task. As to which part this is, we usually have no idea.
Missing Attributes Conversely, some critical attributes can be missing. Mindful of
his parents’ finances, Johnny may be prejudiced against expensive pies. The absence
of attribute price will then make it impossible to induce a good classifier: two
examples, identical in terms of the available attributes, can differ in the values of the
vital “missing” attribute. No wonder that, though identically described, one example
is positive, and the other is negative. When this happens, we say that the training set
is inconsistent. The situation is sometimes difficult to avoid: not only may the expert
be ignorant of the relevance of attribute price; it may be impossible to provide this
attribute’s values, and the attribute thus cannot be used anyway.
Redundant Attributes Somewhat less damaging are attributes that are redundant
in the sense that their values can be obtained from other attributes. If the database
contains a patient’s date-of-birth as well as age, the latter is unnecessary
because it can be calculated by subtracting date-of-birth from today’s date.

Fortunately, redundant attributes are less dangerous than irrelevant or missing ones.
Missing Attribute Values In some applications, the user has no problems identi-
fying the right choice of attributes. The problem is, however, that the value of some
attributes are not known. For instance, the company analyzing the database of its
employees may not know, for each person, the number of children.
Attribute: Value Noise Attribute values and class labels often cannot be trusted on
account of unreliable sources of information, poor measurement devices, typos, the
user’s confusion, and many other reasons. We say that the data suffer from various
kinds of noise.
Stochastic noise is random. For instance, since our body-weight varies during the
day, the reading we get in the morning is different from the one in the evening. A
human error can also play a part: lacking the time to take a patient’s blood pressure,
a negligent nurse simply scribbles down a modification of the previous reading.
By contrast, systematic noise drags all values in the same direction. For instance,
a poorly calibrated thermometer always gives a lower reading than it should. And
something different occurs in the case of arbitrary artifacts; here, the given value
bears no relation to reality such as when an EEG electrode gets loose and, from that
moment on, all subsequent readings will be zero.
Class-Label Noise Class labels suffer from similar problems as attributes. The
labels recommended by an expert may not have been properly recorded; alter-
natively, some examples find themselves in a “gray area” between two classes,
in which event the correct labels are not certain. Both cases represent stochastic
noise, of which the latter may affect negatively only examples from the borderline
region between the two classes. However, class-label noise can also be systematic:
a physician may be reluctant to diagnose a rare disease unless the evidence is
overwhelming—his class labels are then more likely to be negative than positive.
Finally, arbitrary artifacts in class labels are encountered in domains where the
classes are supplied by an automated process that has gone wrong.
Class-label noise can be more dangerous than attribute-value noise. Thus in
the “circles” domain, an example located deep inside the positive region will stay
there even if an attribute’s value is slightly modified; only the borderline example
will suffer from being “sent across the border.” By contrast, class-label noise will
invalidate any example.
• Explain the following types of attributes: irrelevant, redundant, and missing.
Illustrate each of them using the “pies” domain.
1.6 Summary and Historical Remarks 15
Learning:
Training
Concept Classifier
Examples
Application:
Concept
Classifier Example
Label
Fig. 1.7 The training examples are used to induce a classifier. The classifier is then employed to
classify future examples
• What is meant by “inconsistent training set”? What can be the cause? How can it
affect the learning process?
• What kinds of noise do we know? What are their possible sources?
1.6 Summary and Historical Remarks
• Induction from a training set of pre-classified examples is the most deeply studied
machine-learning task.
• Historically, the task is cast as search. One can propose a mechanism that exploits
the well-established search technique of hill climbing defined by an initial state,
final state, interim states, search operators, and evaluation functions.
• Mechanical use of search is not the ultimate solution, though. The rest of the
book will explore more useful techniques.
• Classifier performance is estimated with the help of pre-classified testing data.
The simplest performance criterion is error rate, the percentage of examples
misclassified by the classifier. The baseline scenario is shown in Fig. 1.7.
• Two classifiers that both correctly classify all training examples may differ
significantly in their handling of the testing set.
• Apart from low error rate, some applications require that the classifier provides
the reasons behind the classification.
• The quality of the induced classifier depends on training examples. The quality
of the training examples depends not only on their choice, but also on the
attributes used to describe them. Some attributes are relevant, others irrelevant
or redundant. Quite often, critical attributes are missing.
• The attribute values and class labels may suffer from stochastic noise, systematic
noise, and random artefacts. The value of an attribute in a concrete example may
not be known.
Historical Remarks The idea of casting the machine-learning task as search was
popular in the 1980s and 1990s. While several “founding fathers” came to see things
this way independently of each other, Mitchell [67] is often credited with being the
first to promote the search-based approach; just as influential, however, was the
family of AQ-algorithms proposed by Michalski [59]. The discipline got a major
boost by the collection of papers edited by Michalski et al. [61]. They framed the
mindset of a whole generation.
There is much more to search algorithms. The interested reader is referred to
textbooks of Artificial Intelligence, of which perhaps the most comprehensive is
Russell and Norvig [84] or Coppin [17].
The reader may find it interesting that the question of proper representation
of concepts or classes intrigued philosophers for centuries. Thus John Stuart Mill
[65] explored concepts that are related towhat the next chapter calls probabilistic
0.1 a
2.5 b 1.4 c 2.3 d
2.1 e 0.3 f g h i j k
2.0 0.2 2.0 1.0 3.2
l m
1.5 3.5
Fig. 1.8 Determine the order in which these search states are visited by heuristic search
algorithms. The numbers next to the “boxes” give the values of the evaluation function for the
individual search states
representation; and William Whewel [96] advocated prototypical representations

that are close to the subject of our Chap. 3.
1.7 Solidify Your Knowledge
The exercises are to solidify the acquired knowledge. The suggested thought
experiments will help the reader see this chapter’s ideas in a different light and
provoke independent thinking. Computer assignments will force the readers to pay
attention to seemingly insignificant details they might otherwise overlook.
1.7 Solidify Your Knowledge 17
Exercises
1. In the sliding-tiles puzzle, suggest a better evaluation function than the one used
in the text.
2. Figure 1.8 shows a search tree where each node represents one search state and
is tagged with the value of the evaluation function. In what order will these states
be visited by hill-climbing search?
3. Suppose the evaluation function in the “pies” domain calculates the percentage
of correctly classified training examples. Let the initial state be the expression
describing the second positive example in Table 1.1. Hand-simulate the hill-
climbing search that uses generalization and specialization operators.
4. What is the size of the instance space in a domain where examples are described
by ten boolean attributes? How large is then the space of classifiers?
Give It Some Thought
1. In the “pies” domain, the size of the space of all classifiers is 2108 , provided that
each subset of the instance space can be represented by a distinct classifier. How
much will the search space shrink if we permit only classifiers in the form of
conjunctions of attribute-value pairs?
2. What kind of noise can you think of in the “pies” domain? What can be the source
of this noise? What other issues may render training sets of this kind less than
perfect?
3. Some classifiers behave as black boxes that do not offer much in the way of
explanations. This, for instance, was the case of the “circles” domain. Suggest
examples of domains where black-box classifiers are impractical, and suggest
domains where this limitation does not matter.
4. Consider the data-related difficulties summarized in Sect. 1.5. Which of them are
really serious, and which can perhaps be tolerated?
5. What is the difference between redundant attributes and irrelevant attributes?
6. Take a class that you think is difficult to describe—for instance, the recognition
of a complex biological object (oak tree, ostrich, etc.) or the recognition of a
music genre (rock, folk, jazz, etc.). Suggest the list of attributes to describe the
training examples. Are the values of these attributes easy to obtain? Which of the
problems discussed in this chapter do you expect will complicate the learning
process?
Computer Assignments
1. Write a program implementing hill climbing and apply it to the sliding-tiles

puzzle. Choose appropriate representation for the search states, write a module
that decides whether a state is a final state, and implement the search operators.
Define two or three alternative evaluation functions and observe how each of
them leads to a different sequence of search steps.
2. Write a program that will implement the “growing circles” algorithm from
Sect. 1.3. Create a training set of two-dimensional examples such as those in
Fig. 1.5. The learning program will use the hill-climbing search. The evaluation
function will calculate the percentage of training examples correctly classified by
the classifier. Consider the following search operators: (1) increase/decrease the
radius of the circle, (2) use a different training example as the circle’s center.
3. Write a program that will implement the search for the description of the “pies
that Johnny likes.” Define your own generalization and specialization operators.
The evaluation function will rely on the error rate observed on the training
examples.
Chapter 2
Probabilities: Bayesian Classifiers
The earliest attempts to predict an example’s class based on the known attribute
values go back to well before World War II—prehistory, by the standards of
computer science. Of course, nobody used the term “machine learning,” in those
days, but the goal was essentially the same as the one addressed in this book.
Here is the essence of the solution strategy they used: using the Bayesian
probabilistic theory, calculate for each class the probability of the given object
belonging to it, and then choose the class with the highest value.
2.1 The Single-Attribute Case
Let us start with something so simple as to be unrealistic: a domain where each

example is described with a single attribute. Once we have developed the basic
principles, we will generalize them so that they can be used in more practical
domains.
Probabilities The basics are easily explained using the toy domain from the
previous chapter. The training set consists of twelve pies (Nall D 12), of which
six are positive examples of the given concept (Npos D 6) and six are negative
(Nneg D 6). Assuming that the examples represent faithfully the real situation, the
probability of Johnny liking a randomly picked pie is therefore 50%:
Npos 6
P.pos/ D D D 0:5 (2.1)
Nall 12
Let us now take into consideration one of the attributes, say, filling-size. The
training set contains eight examples with thick filling (Nthick D 8). Out of these, three
are labeled as positive (Nposjthick D 3). This means that the “conditional probability

DOI 10.1007/978-3-319-63913-0_2
20 2 Probabilities: Bayesian Classifiers
Fig. 2.1 The prior

6 Johnny likes Thick Filling
probabilities, P.pos/ D 12
8
and P.thick/ D 12 ; the
conditional probabilities,
P.posjthick/ D 38 and
P.thickjpos/ D 36 ; and the
joint probability,
3
P.likes; thick/ D 12
of an example being positive given that filling-size=thick” is 37.5%—this

is what the relative frequency of positive examples among those with thick filling
implies:
Nposjthick 3
P.posjthick/ D D D 0:375 (2.2)
Nthick 8
Applying Conditional Probability to Classification Importantly, the relative fre-

quency is calculated only for pies with the given attribute value. Among these same
eight pies, five represented the negative class, which means that P.negjthick/ D
5=8 D 0:625. Observing that P.negjthick/ > P.posjthick/, we conclude
that the probability of Johnny disliking a pie with thick filling is greater than the
probability of the opposite case. It thus makes sense for the classifier to label all
examples with filling-size=thick as negative instances of the “pie that
Johnny likes.”
Note that conditional probability, P.posjthick/, is more trustworthy than the
prior probability, P.pos/, because of the additional information that goes into its
calculation. This is only natural. In a DayCare center where the number of boys is
about the same as that of girls, we expect a randomly selected child to be a boy with
P.boy/ D 0:5. But the moment we hear someone call the child Johnny, we increase
this expectation, knowing that it is rare for a girl to have this name. This is why
P.boyjJohnny/ > P.boy/.
Joint Probability Conditional probability should not be confused with joint
probability of two events occurring simultaneously. Be sure to use the right notation:
in joint probability, the terms are separated by commas, P.pos; thick/; in
conditional probability, by a vertical bar, P.posjthick/. For a randomly picked
pie, P.pos; thick/ denotes the probability that the example is positive and its
filling is thick; whereas P.posjthick/ refers to the occurrence of a positive
example among those that have filling-size=thick.
2.1 The Single-Attribute Case 21
A Concrete Example Figure 2.1 illustrates the terms. The rectangle represents
all pies. The positive examples are contained in one circle and those with
filling-size=thick in the other; the intersection contains three instances
that satisfy both conditions; one pie satisfies neither, and is therefore left outside
both circles. The conditional probability, P.posjthick/ D 3=8, is obtained by
dividing the size of the intersection (three) by the size of the circle thick (eight).
The joint probability, P.pos; thick/ D 3=12, is obtained by dividing the size
of the intersection (three) by the size of the entire training set (twelve). The prior
probability of P.pos/ D 6=12 is obtained by dividing the size of the circle pos
(six) with that of the entire training set (twelve).
Obtaining Conditional Probability from Joint Probability The picture con-
vinces us that joint probability can be obtained from prior probability and condi-
tional probability:
3 8 3
P.pos; thick/ D P.posjthick/ P.thick/ D D
8 12 12
3 6 3
P.thick; pos/ D P.thickjpos/ P.pos/ D D
6 12 12
Note that joint probability can never exceed the value of the corresponding
conditional probability: P.pos; thick/ P.posjthick/. This is because
conditional probability is multiplied by prior probability, P.thick/ or P.pos/,
which can never be greater than 1.
Another fact to notice is that P.thick; pos/ D P.pos; thick/ because
both represent the same thing: the probability of thick and pos co-occurring.
Consequently, the left-hand sides of the previous two formulas have to be equal,
which implies the following:
P.posjthick/ P.thick/ D P.thickjpos/ P.pos/
Dividing both sides of this last equation by P.thick/, we obtain the famous
Bayes formula, the foundation for the rest of this chapter:
P.thickjpos/ P.pos/
P.posjthick/ D (2.3)
P.thick/
If we derive the analogous formula for the probability that pies with
filling-size = thick will belong to the negative class, we obtain the
following:
P.thickjneg/ P.neg/
P.negjthick/ D (2.4)
P.thick/
Comparison of the values calculated by these two formulas will tell us which
class, pos of neg, is more probable. Things are simpler than they look: since the
denominator, P.thick/, is the same for both classes, we can just as well ignore it
and simply choose the class for which the numerator is higher.
A Trivial Numeric Example That this formula leads to correct values is illustrated
in Table 2.1 which, for the sake of simplicity, deals with the trivial case where
the examples are described by a single boolean attribute. So simple is this single-
attribute world, actually, that we might easily have obtained P.posjthick/ and
P.negjthick/ directly from the training set, without having to resort to the mighty
Bayes formula—this makes it easy to verify the correctness of the results.
When the examples are described by two or more attributes, the way of
calculating the probabilities is essentially the same, but we need at least one more
trick. This will be introduced in the next section.
• How is the Bayes formula derived from the relation between the conditional and
joint probabilities?
• What makes the Bayes formula so useful? What does it enable us to calculate?
• Can the joint probability, P.x; y/, have a greater value than the conditional
probability, P.xjy/? Under what circumstances is P.xjy/ D P.x; y/?
2.2 Vectors of Discrete Attributes
Let us now proceed to the question how to apply the Bayes formula in
domains where the examples are described by vectors of attributes such as
x D .x1 ; x2 ; : : : ; xn /.
Multiple Classes Many realistic applications have more than two classes, not just
the pos and neg from the “pies” domain. If ci is the label of the i-th class, and if x
is the vector describing the object we want to classify, the Bayes formula acquires
the following form:
P.xjci /P.ci /
P.ci jx/ D
P.x/
2.2 Vectors of Discrete Attributes 23
Table 2.1 Illustrating the principle of Bayesian decision making

Let the training examples be described by a single attribute, filling-size, whose value is
either thick or thin. We want the machine to recognize the positive class (pos). Here are
the eight available training examples:
ex1 ex2 ex3 ex4 ex5 ex6 ex7 ex8

Size thick thick thin thin thin thick thick thick
Class pos pos pos pos neg neg neg neg
The probabilities of the individual attribute values and class labels are obtained by their
relative frequencies. For instance, three out of the eight examples are characterized by
filling-size=thin; therefore, P.thin/ D 3=8.
P.thin/ D 3=8
P.thick/ D 5=8
P.pos/ D 4=8
P.neg/ D 4=8
The conditional probability of a concrete attribute value within a given class is, again,
determined by relative frequency. Our training set yields the following values:
P.thinjpos/ D 2=4
P.thickjpos/ D 2=4
P.thinjneg/ D 1=4
P.thickjneg/ D 3=4
Using these values, the Bayes formula gives the following conditional probabilities:
P.posjthin/ D 2=3
P.posjthick/ D 2=5
P.negjthin/ D 1=3
P.negjthick/ D 3=5
(note that P.posjthin/ C P.negjthin/ D P.posjthick/ C P.negjthick/ D 1)
Based on these results, we conclude that an example with filling-size=thin should

be classified as positive because P.posjthin/ > P.negjthin/. Conversely, an example
with filling-size = thick should be classified as negative because P.negjthick/ >
P.posjthick/.
The denominator being the same for each class, we choose the class that
maximizes the numerator, P.xjci /P.ci /. Here, P.ci / is easy to estimate by the relative
frequency of ci in the training set. As for P.xjci /, however, things are not so simple.
A Vector’s Probability P.xjci / is the probability that a randomly selected repre-

sentative of class ci is described by vector x. Can its value be estimated by relative
frequency? Not really. In the “pies” domain, the size of the instance space was 108
different examples, of which the training set contained twelve. These twelve vectors
were each represented by one training example, while none of the other vectors (the
vast majority!) was represented at all. The relative frequency of x among the six
positive examples was thus either P.xjpos/ D 1=6, when x was among them, or
P.xjpos/ D 0, when it was not. Any x identical to a training example “inherits”
this example’s class label; if the vector is not found in the training set, we have
P.xjci / D 0 for any ci . The numerator in the Bayes formula thus being always
P.xjci /P.ci / D 0, we are unable to choose the most probable class. Evidently, we
will not get very far calculating the probability of an event that occurs only once or
not at all.
This, fortunately, is not the case with the individual attributes. For
instance, shape=circle occurs four times among the positive examples
and twice among the negative, the corresponding probabilities thus being
P.shape D circlejpos/ D 4=6 and P.shape D circlejneg/ D 2=6. If
an attribute can acquire only two or three values, chances are high that each of these
values is represented in the training set more than once, thus offering better grounds
for probability estimates.
Mutually Independent Attributes What is needed is a formula that combines
probabilities of individual attribute values into the probability of the given attribute
vector in the given class: P.xjci /. As long as the attributes are independent of
each other, this is simple. If P.xi jcj / is the probability that the value of the i-th
attribute of an example from class cj is xi , then the probability, P.xjcj /, that a random
representative of cj is described by x D .x1 ; x2 ; : : : ; xn /, is calculated as follows:
Y
n
P.xjcj / D P.xi jcj / (2.5)
iD1
An object will be labeled with cj if this class maximizes the following version of
the Bayes formula’s numerator:
Y
n
P.cj / P.xi jcj / (2.6)
iD1
The Naive Bayes Assumption The reader may complain that the assumption of
mutually independent attributes is rarely justified. Indeed, can the interrelation of
diverse variables ever be avoided? An object’s weight grows with its size, the quality
of health care may be traced to an individual’s income, an object’s color can be
derived from its physical properties. In short, domains where no two attributes are in
any way related to each other are rare. No wonder that the above-described approach
is known under the unflattering name, Naive Bayes.
2.2 Vectors of Discrete Attributes 25
Yet practical experience is not bad at all. True, the violation of the “independence
requirement” renders the probability estimates inaccurate. However, this does not
necessarily make them point to the wrong classes. Remember? x is labeled with the
class that maximizes P.xjci / P.ci /. If the product’s value is 0.8 for one class and
0.2 for the other, then the classifier’s behavior will not change even if the probability
estimates miss the accuracy mark by ten or 20%. And so, while requesting that
the attributes in principle be independent, we will do reasonably well even if they
are not.
When Mutual Independence Cannot Be Assumed This said, we have to ask how
to handle the case where attribute interdependence cannot be ignored. A scientist’s
first instinct may be to suggest more sophisticated ways of estimating P.xjci /. These
do indeed exist, but their complexity grows with the number of attributes, and they
contain terms whose values are hard to determine. The practically minded engineer
doubts that the trouble is justified by the benefits it brings.
A more pragmatic approach will therefore seek to reduce the attribute depen-
dence by appropriate data pre-processing. A good way to start is to get rid
of redundant attributes, those whose values are known to depend on others.
For instance, if the set of attributes contains age, date-of-birth, and
current-date, chances are that Naive Bayes will do better if we use only age.
We can also try to replace two or three attributes by an artificially created
one that combines them. Thus in the “pies” domain, a baker might have
told us that filling-size is not quite independent of crust-size: if
one is thick, the other is thin and vice versa. In this event, we may
benefit from replacing the two attributes with a new one, say, CF-size,
that acquires only two values: thick-crust-and-thin-filling or
thin-crust-and-thick-filling.
In the last resort, if we are prejudiced against advanced methods of multivariate
probability estimates, and if we want to avoid data pre-processing, there is always
the possibility of giving up on Bayesian classifiers altogether, preferring some of the
machine-learning paradigms from the later chapters of this book.
A Numeric Example To get used to the mechanism in which Naive Bayes is
used for classification purposes, the reader may want to go through the example
in Table 2.2. Here the class of a previously unseen pie is established based on
the training set from Table 1.1. The Bayes formula is used, and the attributes are
assumed to be mutually independent.
The procedure is summarized by the pseudocode in Table 2.3.
• Under what circumstances shall we assume that the individual attributes are
mutually independent? What benefit does this assumption bring for the estimates
of P.xjci /?
• Discuss the conflicting aspects of this assumption.
2.3 Probabilities of Rare Events: Exploiting the Expert’s

Intuition
In the first approximation, probability is almost identified with relative frequency:

having observed x thirty times in one hundred trials, we assume that P.x/ D 0:3.
This is how we did it in the previous sections.
Table 2.2 Bayesian classification: examples described by vectors of independent attributes

Suppose we want to apply the Bayesian formula to the training set from Table 1.1 in order to
determine the class of the following object:
x = [shape=square, crust-size=thick, crust-shade=gray

filling-size=thin, filling-shade=white]
There are two classes, pos and neg. The procedure is to calculate the numerator of the Bayes
formula separately for each of them, and then choose the class with the higher value. In the
Qnof representatives: P.pos/ D P.neg/ D 0:5.
training set, each classQhas the same number
n
The remaining terms, iD1 P.xi jpos/ and iD1 P.xi jneg/, are calculated from the following
conditional probabilities:
P(shape=square|pos) D 1=6 P(shape=square|neg) D 2=6

P(crust-size=thick|pos) D 5=6 P(crust-size=thick|neg) D 5=6
P(crust-shade=gray|pos) D 1=6 P(crust-shade=gray|neg) D 2=6
P(filling-size=thin|pos) D 3=6 P(filling-size=thin|neg) D 1=6
P(filling-shade=white|pos) D 1=6 P(filling-shade=white|neg) D 2=6
Based on these values, we obtain the following probabilities:
Y
n
1 5 1 3 1 15
P.xjpos/ D P.xi jpos/ D D 5
iD1
6 6 6 6 6 6
Y
n
2 5 2 1 2 40
P.xjneg/ D P.xi jneg/ D D 5
iD1
6 6 6 6 6 6
Since P.xjpos/ < P.xjneg/, we label x with the negative class.

2.3 Probabilities of Rare Events: Exploiting the Expert’s Intuition 27
Table 2.3 Classification with the Naive-Bayes principle

The example to be classified is described by x D .x1 ; : : : ; xn /.
1. For each xi , and for each class cj , calculate the conditional probability, P.xi jcj /, as the relative
frequency of xi among those training examples that belong to cj .
2. For each class, cj , carry out the following two steps:
i) estimate P.cj / as the relative frequency of this class in the training set;
ii) calculate the conditional probability, P.xjcj /, using the “naive” assumption of mutually
independent attributes:
Y
n
P.xjcj / D P.xi jcj /
iD1
Qn
3. Choose the class with the highest value of P.cj / iD1 P.xi jcj /.
To be fair, though, such estimates can be trusted only when supported by a great
many observations. It is conceivable that a coin flipped four times comes up heads
three times, and yet it will be overhasty to interpret this observation as meaning
that P.heads/ D 0:75; the physics of the experiment suggests that a fair coin
should come up heads 50% of the time. Can this prior expectation help us improve
probability estimates in domains with insufficient numbers of observations?
The answer is, “Yes, we can use the m-estimate.”
The Essence of an m-Estimate Let us illustrate the principle using the case of an
unfair coin where one side comes up somewhat more frequently than the other. In
the absence of any better guidance, the prior expectation of heads is head D 0:5.
An auxiliary parameter, m, helps the engineer tell the class-predicting program how
confident he is in this value, how much the prior expectation can be trusted (higher
m indicating higher confidence).
Let us denote by Nall the number of times the coin was flipped, and by Nheads the
number of times the coin came up heads. The way to combine these values with the
prior expectation and confidence is summarized by the following formula:
Nheads C mheads
Pheads D (2.7)
Nall C m
Note that the formula degenerates to the prior expectation, heads , if Nall D
Nheads D 0. Conversely, it converges to that of relative frequency if Nall and Nheads
are so large as to render the terms mheads and m negligible. Using the values
heads D 0:5 and m D 2, we obtain the following:
Nheads C 2 0:5 Nheads C 1
Pheads D D
Nall C 2 Nall C 2
Illustrating Probability Estimates Table 2.4 shows how the values thus calculated
gradually evolve in the course of five trials. The reader can see that the m-estimate
is for small numbers of experiments more in line with common sense than relative
Table 2.4 For each successive trial, the second row gives the observed outcome; the third, the
relative frequency of heads; the last, the m-estimate of the probability, assuming heads D 0:5 and
mD2
Toss number 1 2 3 4 5
Outcome Heads Heads Tails Heads Tails
Relative frequency 1.00 1.00 0.67 0.75 0.60
m-estimate 0.67 0.75 0.60 0.67 0.57
frequency. Thus after two trials, m-estimate suggests a 0.75 chance of heads,
whereas anybody espousing relative frequency will have to concede that, based on
the two experiments, there is a zero chance that the coin will come up tails. As the
number of trials increases, though, the values returned by m-estimate and relative
frequency tend to converge.
The Impact of the User’s Confidence Let us take a closer look at the effect of
m, the user’s confidence. A lot is revealed if we compare the two different settings
below: m D 100 on the left and m D 1 on the right (in both cases, heads D 0:5).
Nheads C 50 Nheads C 0:5

Nall C 100 Nall C 1
The version with m D 100 allows the prior estimate to be modified only if really
substantial evidence is available (Nheads 50; Nall 100). By contrast, the version
with m D 1 allows the user’s opinion to be controverted with just a few experimental
trials.
Domains with More Than Two Outcomes Although we have used a two-outcome
domain, the formula is applicable also in multi-outcome domains. Rolling a fair die
can result in six different outcomes, and we expect that the probability of seeing,
say, three points is three D 1=6. Using m D 6, we obtain the following:
1
Nthree C mthree Nthree C 6 6 Nthree C 1
Pthree D D D
Nall C m Nall C 6 Nall C 6
Again, if Nall is so high that m D 6 and mthree D 1 can be neglected, the formula
converges to relative frequency: Pthree D NNthree
all
. If we do not want this to happen
prematurely (perhaps because we have high confidence in the prior estimate, three ),
we prevent it by choosing a higher m.
The Limits of m-Estimates We should not forget that the m-estimate is only as
good as the parameters it relies on. If we start from an unrealistic prior estimate, the
result can be disappointing. Suppose that heads D 0:9 and m D 10. Equation (2.7)
then turns into the following:
Nheads C 9
Pheads D
Nall C 10
2.3 Probabilities of Rare Events: Exploiting the Expert’s Intuition 29
When we use this formula to recalculate the values from Table 2.4, we will realize
3C9
that, after five trials, the probability is estimated as Pheads D 5C10 D 12
15
D 0:8,
surely a less plausible value than the one obtained in the case of heads D 0:5 where
we got Pheads D 0:57. The reader is encouraged to verify that the situation will
somewhat improve if we reduce m.
Mathematical Soundness Let us make one last comment. A common understand-
ing in mathematics is that the probabilities of all possible events should sum up to
1: if an experiment canPhave N different outcomes, and if Pi is the probability of
the i-th outcome, then NiD1 Pi D 1. It is easy to verify that Eq. (2.7) satisfies this
condition for any value of m. Suppose we are dealing with the coin-tossing domain
where there are only two possible outcomes. If the prior estimates sum up to 1
(heads C tails D 1), then, given that Nheads C Ntails D Nall , we derive the following:
Nheads Cmheads Ntails Cmtails

Pheads C Ptails = Nall Cm
C Nall Cm
Nheads CNtails Cm.heads Ctails /

= Nall Cm
D1
The interested reader will easily generalize this to any finite number of classes.
Why This May Be Useful In the problem presented in Table 2.5, we want to
classify example x using the Bayesian classifier. To be able to do that, we first need
to calculate the requisite conditional probabilities. Trying to do so for the positive
class, however, we realize that, since the training set is so small, none of the training
examples has crust-shade=gray, the value observed in x. If the probabilities
are estimated by relative frequency, this concrete conditional probability would be
0. As a result, P.xjpos/ D 0, regardless of all the other probabilities. This simply
does not seem right.
The problem disappears if we use m-estimate instead of relative frequency
because the m-estimate is non-zero even if the concrete value has never being
observed in the training set.
• Under what circumstances is relative frequency ill-suited for estimates of discrete
probabilities?
• What is the impact of parameter m in Eq. (2.7)? Under what circumstances will
you prefer large m, and when will you rather go for small m?
• What is the impact of the prior estimate, heads , in Eq. (2.7)? How is the
credibility of m-estimates affected by unrealistic values of heads ?
Table 2.5 An example of one reason for using m-estimates in Bayesian classification
Let us return to the “pies” domain from Table 1.1. Remove from the table the first example, then
use the rest for the calculation of the probabilities.
x = [shape=circle, crust-size=thick, crust-shade=gray

filling-size=thick, filling-shade=dark]
Let us first calculate the probabilities of the individual attribute values:
P(shape=circle|pos) D 3=5
P(crust-size=thick|pos) D 4=5
P(crust-shade=gray|pos) D 0=5
P(filling-size=thick|pos) D 2=5
P(filling-shade=dark|pos) D 3=5
Based on these values, we obtain the following probabilities:
34023
P.xjpos/ D 55
D 0.
We see that the circumstance that none of the five positive examples has crust-shade=gray
causes the corresponding conditional probability to equal 0.
The problem is solved if we calculate the probabilities using the m-estimate. In this case, none
of the conditional probabilities will be 0.
2.4 How to Handle Continuous Attributes
Up till now, we limited our considerations to attributes that assume discrete values,
estimating their probabilities either by relative frequency or by the m-estimate. This,
however, is not enough. In many applications, we encounter attributes (such as
age, price or weight) that acquire values from continuous domains.
Relative frequency is then impractical. While it is easy to establish that the
probability of an engineering student being male is Pmale D 0:7, the probability
that this student’s body weight is 184.5 pounds cannot be specified so readily: the
number of different weight values being infinite, the probability of any one of
them is infinitesimally small. What to do in this case?
Discretizing Continuous Attributes One possibility is to discretize. The simplest
“trick” will split the attribute’s original domain in two; for instance, by replacing
age with the boolean attribute old that is true for age > 60 and false otherwise.
However, at least part of the available information then gets lost: a person may be
old, but we no longer know how old; nor do we know whether one old person is
older than another old person.
2.4 How to Handle Continuous Attributes 31
The loss will be mitigated if we divide the original domain into not two, but
several intervals, say, .0; 10; : : : .90; 100.1 Suppose we get ourselves a separate
bin for each of these, and place a little black ball into the i-th bin for each training
example whose value of age falls into the i-th interval.
Having done so, we may reach the situation depicted in Fig. 2.2. The upper part
shows the bins, side by side, and the bottom part shows apiecewise constant function
Fig. 2.2 A simple

. .. .
. . ... ..... ...... .... .. ... .. . . .
discretization method that
represents each subinterval by
a separate bin. The bottom
x
chart plots the histogram over b1 b2 b3 b b5 b6 b b8
4 7
the individual subintervals
p(x)
created in the following manner: if N is the size of the training set, and Ni is the
number of balls in the i-th bin, then the function’s value in the i-th interval is Ni =N—
the relative frequency of the balls in the i-th bin. Since the area under the function is
†Ni
N
D 1, we have a mechanism to estimate the probability not of a concrete value
of age, but rather of this value falling into the given interval.
Probability Density Function If the step-function thus constructed seems too
crude, we may fine-tune it by dividing the original domain into shorter—and thus
more numerous—intervals, provided that the number of balls in each bin is sufficient
for reliable probability estimates. If the training set is infinitely large, we can,
theoretically speaking, keep reducing the lengths of the intervals until they become
infinitesimally small. The result of the bin-filling exercise will then no longer be
a step-function, but rather a continuous function, p.x/, such as the one in Fig. 2.3.
Its interpretation follows from the way it has been created: a high value of p.x/
indicates that there are many examples with age close to x; conversely, a low
value of p.x/ tells us that age in the vicinity of x is rare. This is why we call
p.x/ a probability density function, often avoiding this mouthful by preferring the
acronym pdf.
1
We assume here that 100 is the maximum value observed in the training set. Alternatively, our
background knowledge may inform us that the given attribute’s value cannot exceed 100.
Fig. 2.3 When using the pdf,

p(x)
we identify the probability of Gaussian "bell" function
x 2 Œa; b with the relative
size of the area below the
corresponding section of the
pdf
p a b age
Let us be careful about notation. The discrete probability of x is indicated by an

uppercase letter, P.x/. The pdf at x is denoted by a lowercase letter, p.x/. And if we
want to point out that the pdf has been created exclusively from examples belonging
to class ci , we do so by using a subscript, pci .x/.
Bayes Formula for Continuous Attributes The good thing about the pdf is that
it makes it possible to employ the Bayes formula even in the case of continuous
attributes. We only replace the conditional probability P.xjci / with pci .x/, and P.x/
with p.x/. Let us begin with the trivial case where the object to be classified is
described by a single continuous attribute, x. The Bayes formula then assumes the
following form:
pci .x/ P.ci /

P.ci j x/ D (2.8)
p.x/
Here, P.ci / is estimated by the relative frequency of ci in the training set, p.x/ is
the pdf created from all training examples, and pci .x/ is the pdf created from those
training examples that belong to ci .
Again, the denominator can be ignored because it has the same value for any
class. The classifier simply calculates, separately for each class, the value of the
numerator, pci .x/ P.ci /, and then labels the object with the class for which the
product is maximized.
Naive Bayes Revisited When facing the more realistic case where the examples
are described by vectors of attributes, we will avail ourselves of the same “trick”
as before: the assumption that all attributes are mutually independent. Suppose we
encounter an example described by x D .x1 ; : : : ; xn /. The pdf at x is approximated
by the product along the individual attributes:
pcj .x/ D …niD1 pcj .xi / (2.9)
A statistician will be able to suggest formulas that are theoretically sounder;

however, higher sophistication often fails to give satisfaction. For one thing, we
may commit the sin of accurate calculations with imprecise numbers. Besides, the
more complicated the technique, the higher the danger it will be applied incorrectly.
2.5 Gaussian “Bell” Function: A Standard pdf 33
• What is the probability density function, pdf, and how does it help us in the
context of Bayesian classification?
• Explain the discretization mechanism that helped us arrive at an informal
definition of a pdf.
• How does the Bayes formula change in domains with continuous attributes? How
do we estimate the values of the individual terms?
2.5 Gaussian “Bell” Function: A Standard pdf
One way to approximate a pdf is to employ the discretization technique from the
previous section. Alternatively, we can capitalize on standardized models known to
be applicable to many realistic situations. Perhaps the most popular among them is
the gaussian function, named after the great German mathematician.
The Shape and the Formula Describing It The curve in Fig. 2.3 is an example; its
shape betrays why many people call it a “bell function.” The maximum is reached
at the mean, x D , and the curve slopes down gracefully with the growing distance
of x from . It is reasonable to expect that this is a good model of the pdf of such
variables as the body temperature where the density peaks around x D 99:7 degrees
Fahrenheit.
Expressed mathematically, the gaussian function is defined by the following
formula where e is the base of the natural logarithm:
.x/2
p.x/ D k e 2 2 (2.10)
Parameters Note that the greater the difference between x and , the greater the
exponent’s numerator, and thus the smaller the value of p.x/ because the exponent
is negative. The reason the numerator is squared, .x /2 , is to make sure that the
value slopes down with the same angle on both sides of the mean, ; the curve
is symmetric. How steep the slope is depends on 2 , a parameter called variance.
Greater variance means smaller sensitivity to the difference between x and , and
thus a “flatter” bell curve; conversely, smaller variance defines a narrower bell curve.
The task for the coefficient k is to make the area under the bell function equal to
1 as required by the theory of probability. It would be relatively easy to prove that
this happens when k is determined by the following formula:
1
kD p (2.11)
2 2
Setting the Parameter Values To be able to use this model when approximating
pci .x/ in a concrete application, we only need to estimate the values of its
parameters, and 2 . This is easy. Suppose that class ci has m representatives
among the training examples. If xi is the value of the given attribute in the i-
th example, then the mean and variance, respectively, are calculated using the
following formulas:
1X
m
D xi (2.12)
m iD1
1 X
m
2 D .xi /2 (2.13)
m 1 iD1
In plain English, the gaussian center, , is obtained as the arithmetic average of

the values observed in the training examples, and the variance is obtained as the
average of the squared differences between xi and . Note that, when calculating
variance, we divide the sum by m 1, and not by m, as we might expect. The
intention is to compensate for the fact that itself is only an estimate. The variance
should therefore be somewhat higher than what it would be if we divided by m.
Of course, this matters only if the training set is small: for large m, the difference
between m and m 1 is negligible.
• Give an example of a continuous variable whose pdf can be expected to follow
the gaussian distribution.
• What parameters define the bell function? How can we establish their values
using the training set?
• How—and why—do we normalize the bell function?
2.6 Approximating PDFs with Sets of Gaussians
While the bell function represents a good mechanism to approximate the pdf in
many realistic domains, it is not a panacea. Some variables simply do not behave
that way. Just consider the distribution of body-weight in a group that mixes
grade-school children with their parents. If we create the pdf using the discretization
method, we will observe two peaks: one for the kids, and the other for the grown-
2.6 Approximating PDFs with Sets of Gaussians 35
ups. There may be three peaks if it turns out that body-weight of fathers is
distributed around a higher mean than that of the mothers. And the number of peaks
can be higher still if the families come from diverse ethnic groups.
Combining Gaussian Functions No doubt, a single bell function would misrep-
resent the situation. But what if we combine two or more of them? If we knew
the diverse sources of the examples, we might create a separate gaussian for each
source, and then superimpose the bell functions on each other. Would this solve our
problem?
The honest answer is, “yes, in this ideal case.” In reality, though, prior know-
ledge about diverse sources is rarely available. A better solution will divide the
body-weight values into great many random groups. In the extreme, we may
even go as far as to make each example a “group” of its own, and then identify a
gaussian center with this example’s body-weight, thus obtaining m bell functions
(for m examples).
The Formula to Combine Them Suppose we want to approximate the pdf of a
continuous attribute, x. If we denote by i the value of x in the i-th example, then
the pdf is approximated by the following sum of m functions:
.xi /2

p.x/ D k †m
iD1 e
2 2 (2.14)
As before, the normalization constant, k, is here to make sure that the area under
the curve is 1. This is achieved when k is calculated as follows:
1
kD p (2.15)
m 2
From mathematics, we know that if m is sufficiently high, Eq. (2.14) approxi-

mates the pdf with almost arbitrary accuracy.
Illustrating the Point Figure 2.4 illustrates the approach using a training set
consisting of m D 3 examples, the values of attribute x being x1 D 0:4; x2 D 0:5
and x3 D 0:7. The upper three charts show three bell functions, each centered at
one of these points, the variance always being 2 D 1. The bottom chart shows the
composed pdf created by putting together Eqs. (2.14) and (2.15), using the means,
1 D 0:4, 2 D 0:5, and 3 D 0:7, and 2 D 1:
1 .x0:4/2 .x0:5/2 .x0:7/2

p.x/ D p Œ e 2 C e 2 C e 2
3 2
The Impact of Concrete Parameter Values The practical utility of the pdf thus
obtained (its success when used in the Bayes formula) depends on the choice of 2 .
In Fig. 2.4, we used 2 D 1, but there is no guarantee that this will work in any
future application. To be able to adjust it properly, we need to understand how it
affects the shape of the composite pdf.
Inspecting the gaussian formula, we realize that the choice of a very small value
of 2 causes great sensitivity to the difference between x and i ; the individual bell
functions will be “narrow,” and the resulting pdf will be marked by steep peaks
separated by extended “valleys.” Conversely, the consequence of a high 2 will be
an almost flat pdf. Seeking a compromise between the two extremes, we will do well
if we make 2 dependent on the distances between examples.
The simplest solution will use 2 D max min , where max and min
are the maximum and minimum values of i , respectively. If you think this too
crude, you may consider normalizing the difference by the number of examples:
2 D .max min /=m. Large training sets (with high m) will then lead to smaller
variations that will narrow the contributing gaussians. Finally, in some domains we
might argue that each of the contributing bell functions should have a variance of its
own, proportional to the distance from the center of the nearest other bell function.
In this case, however, we are no longer allowed to set the value of k by Eq. (2.15).
A Numeric Example The example in Table 2.6 illustrates the whole procedure on
a concrete case of a small training set and a vector to be classified. The reader is
encouraged to go through all its details to get used to the way the formulas are put
together. See also the illustration in Fig. 2.5.
When There Are Too Many Examples For a training set of realistic size, it is
impractical to identify each training example with one gaussian centers; nor is it
necessary. More often than not, the examples are grouped in clusters that can be
detected by cluster analysis techniques—see Chap. 14. Once the clusters have been
found, we identify the gaussian centers with the centroids of the clusters.
• Under what circumstances is the gaussian function a poor model of the pdf ?
• Why does the composite pdf have to be normalized by k?
• How do we establish the centers and variances of the individual bell functions?
• Bayesian classifiers calculate the product P.xjci /P.ci / separately for each class,
ci , and then label the example, x, with the class where this product has the highest
value.
• The main problem is how to calculate the probability, P.xjci /. Most of the time,
the job is simplified by making the assumption that Qthe individual attributes are
mutually independent, in which case P.xjci / D njD1 P.xj jci /, where n is the
number of attributes.
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 2.4 Composing the pdf from three examples with the following values of attribute x: 1 D
0:4; 2 D 0:5, and 3 D 0:7. The upper three charts show the contributing gaussians; the bottom
chart, the composition. The variance is 2 D 1
• The so-called m-estimate makes it possible to take advantage of a user’s

estimate of an event’s probability. This comes handy in domains with insufficient
experimental evidence, where relative frequency cannot be relied on.
• In domains with continuous attributes, the role of the discrete probability, P.xjci /,
is taken over by pci .x/, the probability density function, pdf, but otherwise the
Table 2.6 Using Naive Bayes in domains with three continuous attributes
Suppose we are given a training set that consists of the following six examples, ex1 , : : :, ex6 ,
each described by three continuous attributes, at1 , at2 , . . . at3 :
Example at1 at2 at3 Class

ex1 3:2 2:1 2:1 pos
ex2 5:2 6:1 7:5 pos
ex3 8:5 1:3 0:5 pos
ex4 2:3 5:4 2:45 neg
ex5 6:2 3:1 4:4 neg
ex6 1:3 6:0 3:35 neg
Using the Bayes formula, we are to find the most probable class of x D .9; 2:6; 3:3/.
Our strategy is to evaluate ppos .x/ P.pos/ and pneg .x/ P.neg/. Observing that P.pos/ D
P.neg/, we simply label x with pos if ppos .x/ > pneg .x/ and with neg otherwise. When
constructing the pdf, we rely on the independent-attributes assumption. In accordance with
Sect. 2.2, we have:
ppos .x/ D ppos .at1 / ppos .at2 / ppos .at3 / and

pneg .x/ D pneg .at1 / pneg .at2 / pneg .at3 /
The terms on the right-hand

p sides are obtained by Eq. (2.14), in which we use 2 D 1; m D 3,
and, therefore, k D 1= .2/3 . Thus for the first of these terms, we get the following:
1 2 2 2
ppos .at1 / D p Œe0:5.at1 3:2/ C e0:5.at1 5:2/ C e0:5.at1 8:5/
3 2
Note that the values of at1 in the positive examples are 1 D 3:2; 2 D 5:2, and 3 D 8:5,
respectively—see the exponents in the expression. The functions for the remaining five terms,
obtained similarly, are plotted in the rightmost column of Fig. 2.5.
Substituting into these equations the coordinates of x, namely at1 D 9; at2 D 3:6, and at3 D 3:3,
will give us the following:
ppos .x/ D 0:0561 0:0835 0:0322 D 0:00015

pneg .x/ D 0:0023 0:0575 0:1423 D 0:00001
Observing that ppos .x/ > pneg .x/, we label x with the class pos.
Fig. 2.5 Composing the pdf ’s separately for the positive and negative class (with 2 D 1). Each
row represents one attribute, and each of the left three columns represents one example. The
rightmost column shows the composed pdf ’s
procedure is the same: the example is labeled with the class that maximizes the
product, pci .x/P.ci /.
• The concrete shape of the pdf is approximated by discretization, by the use of
standardized pdf s, or by the sum of gaussians.
• The estimates of probabilities are far from perfect, but the results are often
satisfactory even when rigorous theoretical assumptions are not satisfied.
Historical Remarks The first papers to use Bayesian decision theory for classifi-
cation purposes were Neyman and Pearson [72] and [25], but the paradigm gained
momentum only with the advent of the computer, when it was advocated by Chow
[14]. The first to use the assumption of independent attributes was Good [33]. The
idea of approximating pdf ’s by the sum of bell functions comes from Parzen [74].
When provided with perfect information about the probabilities, the Bayesian
classifier is guaranteed to provide the best possible classification accuracy. This
is why it is sometimes used as a reference to which the performance of other
approaches is compared.
Exercises
1. A coin tossed three times came up heads, tails, and tails, respectively. Calculate
the m-estimate for these outcomes, using m D 3 and heads D tails D 0:5.
2. Suppose you have the following training examples, described by three attributes,
x1 ; x2 ; x3 , and labeled by classes c1 and c2 .
x1 x2 x3 Class
2.1 0.2 3.0 c1
3.3 1.0 2.9 c1
2.7 1.2 3.4 c1
0.5 5.3 0.0 c2
1.5 4.7 0.5 c2
Using these data, do the following:

(a) Assuming that the attributes are mutually independent, approximate the
following probability density functions: pc1 .x/; pc2 .x/; p.x/. Hint: use the
idea of superimposed bell functions.
(b) Using the pdf ’s from the previous step, decide whether x D Œ1:4; 3:3; 3:0
should belong to c1 or c2 .
1. How would you apply the m-estimate in a domain with three possible outcomes,
ŒA; B; C, each with the same prior probability estimate, A D B D C D 1=3?
What if you trust your expectations of A but are not so sure about B and C? Is
there a way to reflect this circumstance in the value of the parameter m?
2. Suggest the circumstances under which the accuracy of probability estimates will
benefit from the assumption that attributes are mutually independent. Explain the
advantages and disadvantages.
3. How would you calculate the probabilities of the output classes in a domain
where some attributes are boolean, others discrete, and yet others continuous?
Discuss the possibilities of combining different approaches.
1. Machine learning researchers often test their algorithms using publicly available
benchmark domains. A large repository of such domains can be found at the
following address: www.ics.uci.edu/~mlearn/MLRepository.html. Take a look at
these data and see how they differ in the numbers of attributes, types of attributes,
sizes and so on.
2. Write a computer program that will use the Bayes formula to calculate the class
probabilities in a domain where all attributes are discrete. Apply this program to
our “pies” domain.
3. For the case of continuous attributes, write a computer program that accepts the
training examples in the form of a table such as the one in Exercise 3 above.
Based on these, the program approximates the pdf s, and then uses them to
determine the class labels of future examples.
4. Apply this program to a few benchmark domains from the UCI repository
(choose from those where all attributes are continuous) and observe that the
program succeeds in some domains better than in others.
Chapter 3
Similarities: Nearest-Neighbor Classifiers
Two plants that look very much alike probably represent the same species; likewise,
it is quite common that patients complaining of similar symptoms suffer from
the same disease. In short, similar objects often belong to the same class—an
observation that forms the basis of a popular approach to classification: when asked
to determine the class of object x, find the training example most similar to it. Then
label x with this example’s class.
The chapter explains how to evaluate example-to-example similarities, presents
concrete mechanisms that use these similarities for classification purposes, com-
pares the performance of this approach with that of the Bayesian classifier, and
introduces methods to overcome some inherent weaknesses.
3.1 The k-Nearest-Neighbor Rule
How do we establish that an object is more similar to x than to y? Some may doubt
that this is at all possible. Is a giraffe more similar to a horse than to a zebra?
Questions of this kind raise suspicion. Too many arbitrary and subjective factors
are involved in answering them.
Similarity of Attribute Vectors The machine-learning task, as formulated in this
book, reduces these objections to a minimum. Rather than real objects, the classifier
will compare their attribute-based descriptions—hardly an insurmountable problem.
Thus in the toy domain from Chap. 1, the similarity of two pies can be established
by counting the attributes in which they differ: the fewer the differences, the greater
the similarity. The first row in Table 3.1 gives the attribute values of object x. For
each of the twelve training examples that follow, the rightmost column specifies the
number of differences between the given example and x. The smallest value being
found in the case of ex5 , we conclude that this is the training example most similar
to x, and this suggests that we should label x with pos, the class of ex5 .

DOI 10.1007/978-3-319-63913-0_3
44 3 Similarities: Nearest-Neighbor Classifiers
Table 3.1 Counting the numbers of differences between pairs of discrete-attribute vectors
Crust Filling
Example Shape Size Shade Size Shade Class # differences
x Square Thick Gray Thin White ? –
ex1 Circle Thick Gray Thick Dark pos 3
ex2 Circle Thick White Thick Dark pos 4
ex3 Triangle Thick Dark Thick Gray pos 4
ex4 Circle Thin White Thin Dark pos 4
ex5 Square Thick Dark Thin White pos 1
ex6 Circle Thick White Thin Dark pos 3
ex7 Circle Thick Gray Thick White neg 2
ex8 Square Thick White Thick Gray neg 3
ex9 Triangle Thin Gray Thin Dark neg 3
ex10 Circle Thick Dark Thick White neg 3
ex11 Square Thick White Thick Dark neg 3
ex12 Triangle Thick White Thick Gray neg 4
Of the 12 training examples, ex5 is the one most similar to x
Table 3.2 The simplest version of the k-NN classifier

Suppose we have a mechanism to evaluate the similarly between attribute vectors. Let x denote
the object whose class we want to determine.
1. Among the training examples, identify the k nearest neighbors of x (examples most similar
to x).
2. Let ci be the class most frequently found among these k nearest neighbors.
3. Label x with ci .
Dealing with continuous attributes is just as simple. The fact that each example
can be represented by a point in an n-dimensional space makes it possible to
calculate the geometric distance between any pair of examples, for instance, by
the Euclidean distance (Sect. 3.2 will have more to say about how to measure
distance between vectors). And again, the closer to each other the examples are
in the instance space, the greater their mutual similarity. This, by the way, is
how the nearest-neighbor classifier got its name: the training example with the
smallest distance from x in the instance space is, geometrically speaking, x’s nearest
neighbor.
From a Single Neighbor to k Neighbors In noisy domains, the testimony of the
nearest neighbor cannot be trusted. What if this single specimen’s class label is
incorrect? A more robust approach identifies not one, but several nearest neighbors,
and then lets them vote. This is the essence of the so-called k-NN classifier, where
k is the number of the voting neighbors—usually a user-specified parameter. The
pseudocode in Table 3.2 summarizes the algorithm.
3.1 The k-Nearest-Neighbor Rule 45
Note that, in a two-class domain, k should be an odd number so as to prevent ties.

For instance, a 4-NN classifier might face a situation where the number of positive
neighbors is the same as the number of negative neighbors. This will not happen to
a 5-NN classifier.
As for domains that have more than two classes, using an odd number of nearest
neighbors does not prevent ties. For instance, the 7-NN classifier can realize that
three neighbors belong to class C1 , three neighbors belong to class C2 , and one
neighbor belongs to class C3 . The engineer designing the classifier then needs to
define a mechanism to choose between C1 and C2 .
An Illustration Certain “behavioral aspects” of this paradigm can be made obvious
with the help of a fictitious domain where the examples are described by two
numeric attributes, a situation easy to visualize. Figure 3.1 shows several positive
and negative training examples, and also some objects (the big black dots) whose
classes the k-NN classifier is to determine. The reader can see that objects 1 and 2
are surrounded by examples from the same class, and their classification is therefore
straightforward. On the other hand, object 3 is located in the “no man’s land”
between the positive and negative regions, so that even a small amount of attribute
noise can send it to either side. The classification of such borderline examples is
unreliable.
In the right-hand part of the picture, object 4 finds itself deep in the positive
region, but class noise has mislabeled its nearest neighbor in the training set as
negative. Whereas the 1-NN classifier will go wrong, here, the 3-NN classifier will
give the correct answer because the other two neighbors, which are positive, will
outvote the single negative neighbor.
y _ y _
_ _
+ __ _ + __ _
+ +
+ _ + _
1
+
3 2 _4 +
+ _ + _
+ __ + __
+ + + +
+ +
x x
Fig. 3.1 Object 3, finding itself in the borderline region, is hard to classify. In the noisy domain
shown in the right-hand part, the 1-NN classifier will misclassify object 4, but the mistake is
corrected if the 3-NN classifier is used
To make sure you understand this topic, try to answer the following questions. If
you have problems, return to the corresponding place in the preceding text.
• How can we measure example-to-example similarity in domains where all
attributes are discrete, and how in domains where they are all continuous?
• Under what circumstances will the k-NN classifier (with k > 1) outperform the
1-NN classifier and why?
• Explain why, in 2-class domains, the k in the k-NN classifier should be an odd
number.
• How does attribute noise affect the classification of a “borderline example”?
What is the impact of class-label noise?
3.2 Measuring Similarity
As indicated, a natural way to find the nearest neighbors of object x is to compare

the geometrical distances of the individual training examples from x. Figure 3.1
illustrated this principleusing a domain so simple that the distances could be
Fig. 3.2 The Euclidean

distance between two points y2
in a two-dimensional space is
equal to the length of the
triangle’s hypotenuse
y1
x1 x2
measured by a ruler. Yet the ruler will hardly be our tool of choice if the examples are
described by more than two attributes. What we need is an expression to calculate
the similarity based on attribute values.
Euclidean Distance In a plane, the geometric distance between two points, x D
.x1 ; x2 / and y D .y1 ; y2 /, is obtained
p with the help of the pythagorean theorem as
indicated in Fig. 3.2: d.x; y/ D .x1 y1 /2 C .x2 y2 /2 . This formula is easy to
generalize to a domain with n continuous attributes where the Euclidean distance
between x D .x1 ; : : : ; xn / and y D .y1 ; : : : ; yn / is defined as follows:
q
dE .x; y/ D †niD1 .xi yi /2 (3.1)
3.2 Measuring Similarity 47
Table 3.3 Using the nearest-neighbor principle in a 3-dimensional Euclidean space

Using the following training set of four examples described by three numeric attributes,
determine the class of object x D Œ2; 4; 2.
Distance between
exi and Œ2; 4; 2
p p
ex1 fŒ1; 3; 1; posg .2 1/2 C .4 3/2 C .2 1/2 D 3
p p
ex2 fŒ3; 5; 2; posg .2 3/2 C .4 5/2 C .2 2/2 D 2
p p
ex3 fŒ3; 2; 2; negg .2 3/2 C .4 2/2 C .2 2/2 D 5
p p
ex4 fŒ5; 2; 3; negg .2 5/2 C .4 2/2 C .2 3/2 D 4
Calculating the Euclidean distances between x and the training examples, we realize that x’s
nearest neighbor is ex2 . Its label being pos, the 1-NN classifier returns the positive label.
The same result is obtained by the 3-NN classifier because two of x’s three nearest neighbors
(ex1 and ex2 ) are positive, and only one (ex4 ), is negative.
The way this metric is used in the context of k-NN classifiers is illustrated
in Table 3.3 where the training set consists of four examples described by three
numeric attributes.
A More General Formulation The reader can see that the term under the square-
root symbol is a sum of the squared distances along corresponding attributes.1 This
observation is mathematically expressed as follows:
q
dM .x; y/ D †niD1 d.xi ; yi / (3.2)
This is how the distance is usually calculated in the case of vectors in which
discrete and continuous attributes are mixed (in the symbol, dM .x; y/, this is
indicated by the subscript, M). For instance, we can use d.xi ; yi / D .xi yi /2 for
continuous attributes, whereas for discrete attributes, we put d.xi ; yi / D 0 if xi D yi
and d.xi ; yi / D 1 if xi ¤ yi .
Note that if all attributes are continuous, the formula is identical to Euclidean
distance; and if the attributes are all discrete, the formula simply specifies the
number of attributes in which the two vectors differ. In purely Boolean domains,
where for any attribute only the values true or false are permitted (let us abbreviate
these values as t and f , respectively), this latter case is called Hamming distance,
dH . For instance, the Hamming distance between the vectors x D .t; t; f ; f / and
y D .t; f ; t; f / is dH .x; y/ D 2. In general, however, Eq. (3.2) is meant to be
employed in domains where the examples are described by a mixture of discrete
and continuous attributes.
1
One benefit of these differences being squared, and thus guaranteed to be positive, is that this
prevents negative differences, xi yi < 0, to be subtracted from positive differences, xi yi > 0.
Attribute-to-Attribute Distances Can Be Misleading We must be careful not to

apply Formula (3.2) mechanically, ignoring the specific aspects of the given domain.
Let us briefly discuss two circumstances that make it is easy to go wrong.
Suppose our examples are described by three attributes, size, price, and
season. Of these, the first two are obviously continuous, and the last, discrete. If
x D .2; 1:5; summer/ and y D .1; 0:5; winter/, then Eq. (3.2) gives the following
distance between the two:
p p
dM .x; y/ D .2 1/2 C .1:5 0:5/2 C 1 D 3
Let us first consider the third attribute: summer being different from winter,
our earlier considerations lead us to establish that d.summer; winter/ D 1.
In reality, though, the difference between summer and winter is sometimes
deemed greater than the difference between, say, summer and fall which are
“neighboring seasons.” Another line of reasoning, however, may convince us that
spring and fall are more similar to each other than summer and winter—
at least as far as weather is concerned. We can see that the two values, 0 and 1,
will clearly not suffice, here. Intermediate values should perhaps be considered,
the concrete choice depending on the specific needs of the given application. The
engineer who does not pay attention to factors of this kind may fail to achieve good
results.
Mixing continuous and discrete attributes can be risky in another way. A
thoughtless application of Eq. (3.2) can result in a situation where the difference
between two sizes (e.g., size1 D 1 and size1 D 12, which means that
d.size1 ; size2 / D 112 D 121) can totally dominate the difference between two
seasons (which, in the baseline version, could not exceed 1). This observation is
closely related to the problem of scaling—discussed in the next section.
Distances in General The reader is beginning to understand that the issue of
similarity is far from trivial. Apart from Eq. (3.2), quite a few other formulas have
been suggested, some of them fairly sophisticated.2 While it is good to know they
exist, we will not examine them here because we do not want to digress too far from
our main topic. Suffice it so say that any distance metric has to satisfy the following
requirements:
1. the distance must never be negative;
2. the distance between two identical vectors, x and y, is zero;
3. the distance from x to y is the same as the distance from y to x;
4. the metric must satisfy the triangular inequality: d.x; y/ C d.y; z/ d.x; z/.
2
Among these, perhaps the best-known are the polar distance, the Minkowski metric, and the
Mahalanobis distance.
3.3 Irrelevant Attributes and Scaling Problems 49
• What is Euclidean distance, and what is Hamming distance? In what domains
can they be used? How is distance related to similarity?
• Write down the distance formula for domains where examples are described
by a mixture of continuous and discrete attributes. Discuss the difficulties that
complicate its straightforward application in practice.
• What fundamental properties have to be satisfied by any distance metric?
3.3 Irrelevant Attributes and Scaling Problems
By now, the reader understands the principles of the k-NN classifier well enough
to be able to write a computer program implementing the tool. This is not enough,
though. If applied mechanically, the software may disappoint. It is necessary to
understand why.
The rock-bottom of the nearest-neighbor paradigm is that, “objects are similar
if the geometric distance between the vectors describing them is small.” This
said, we must be careful not to let this principle degenerate into a dogma. In
certain situations, the geometric distance can be misleading. Two of them are quite
common.
Irrelevant Attributes It would be a mistake to think that all attributes are created
equal. They are not. In the context of machine learning, some are irrelevant in the
sense that their values have nothing to do with the given example’s class. But they
do affect the geometric distance between vectors.
A simple illustration will clarify the point. In the training set from Fig. 3.3, the
examples are characterized by two numeric attributes: body-temperature (the
horizontal axis) and shoe-size (the vertical axis). The black dot stands for the
object that the k-NN classifier is expected to label as healthy (pos) or sick (neg).
As expected, all positive examples find themselves in the shaded area delimited
by two critical points along the “horizontal” attribute: temperatures exceeding
the maximum indicate fever; those below the minimum, hypothermia. As for the
“vertical” attribute, though, we see that the positive and negative examples alike
are distributed along the entire range, show-size being unable to betray anything
about a person’s health. The object we want to classify is in the highlighted region,
and common sense requires that it should be labeled as positive—despite the fact
that its nearest neighbor happens to be negative.
The Lesson If only the first pattribute is used, the Euclidean distance between the
two examples is dE .x; y/ D .x1 y1 /2 D p jx1 y1 j. If both attributes are used,
the Euclidean distance will be dE .x; y/ D .x1 y1 /2 C .x2 y2 /2 : If the second
Fig. 3.3 The “vertical”

attribute is irrelevant for shoe + _
classification, and yet it size _
affects the geometric _
distances between examples.
Object 1 is positive, even _
1 ._ _
though its nearest neighbor in
the 2-dimensional space is _
negative _ + _
+
body−temperature
attribute is irrelevant, then the term .x2 y2 /2 is superfluous—and yet it affects,

adversely, k-NN’s notion of similarity! This is what occurred in Fig. 3.3, and this is
why object 1 was misclassified.
How much damage is caused by irrelevant attributes depends on how many of
them are used to describe the examples. In a domain with hundreds of attributes, of
which only one is irrelevant, there is no need to panic: one lonely culprit is unlikely
to distort the value of dE .x; y/ in any meaningful way. But things will change as the
percentage of irrelevant attributes grows. If the vast majority of the attributes have
nothing to do with the class we want to recognize, then the geometric distance will
become almost meaningless, and the classifier’s performance will be dismal.
The Scales of Attribute Values Suppose we want to evaluate the similarity of two
examples, x D .t; 0:2; 254/ and y D .f ; 0:1; 194/, described by three attributes, of
which the first is boolean, the second is continuous with values from interval Œ0I 1,
and the third is continuous with values from interval Œ0I 1000. Using Eq. (3.2), the
reader will find it easy to calculate the distance between x and y, obtaining the
following:
p
dM .x; y/ D .1 0/2 C .0:2 0:1/2 C .254 194/2
Inspecting this expression, we notice that the third attribute completely domi-
nates, reducing the other two to virtual insignificance. No matter how we modify
their values within their ranges, the distance, dM .x; y/, will hardly change. Fortu-
nately, the situation is easy to rectify. If we divide, in the training set, all values of
the third attribute by 1000, thus “squeezing” its range to Œ0I 1, the impacts of the
attributes will become more balanced. We can see that the scales of the attribute
values can radically affect the k-NN classifier’s behavior.
Another Aspect of Attribute Scaling There is more to it. Consider the following
two training examples, ex1 and ex2 , and the object x whose class we want to
determine:
ex1 D Œ.10; 10/; pos/
ex2 D Œ.20; 0/; neg/
x D .32; 20/
3.3 Irrelevant Attributes and Scaling Problems 51
p p
The distances are dM .x; ex1 / D 584 and dM .x; ex2 / D 544. The latter
being smaller, the 1-NN classifier will label x as neg. Suppose, however, that the
second attribute expresses temperature, and does so in centigrades. If we decide
to use Fahrenheits instead, the three vectors will change as follows:
ex1 D Œ.10; 50/; pos/
ex2 D Œ.20; 32/; neg/
x D .32; 68/
p
p Recalculating the distances, we obtain dM .x; ex1 / D 808 and dM .x; ex2 / D
1440. This time, it is the first distance that is smaller, and 1-NN will therefore
classify x as positive. This seems a bit silly. The examples are still the same, except
that we chose different units for temperature; and yet the classifier’s verdict has
changed.
Normalizing Attribute Scales One way out of this trouble is to normalize the
attributes: to re-scale them in a way that makes all values fall into the same interval,
Œ0I 1. From the several mechanisms that have been used to this end, perhaps the
simplest is the one that first identifies, for the given attribute, its maximum (MAX)
and minimum (MIN), and then replaces each value, x, of this attribute using the
following formula:
x MIN
xD (3.3)
MAX MIN
A simple illustration will show us how this works. Suppose that, in the training
set consisting of five examples, a given attribute acquires the following values,
respectively:
Œ7; 4; 25; 5; 10
We see that MIN D 5 and MAX D 25. Subtracting MIN from each of the
values, we obtain the following:
Œ12; 9; 30; 0; 15
The reader can see that the “new minimum” is 0, and the “new maximum” is
MAX MIN D 25 .5/ D 30. Dividing the obtained values by MAX MIN, we
obtain a situation where all the values fall into Œ0I 1:
Œ0:4; 0:3; 1; 0; 0:5
One Potential Weakness of Normalization Normalization reduces error rate in

many practical applications, especially if the scales of the original attributes vary
significantly. The downside is that the description of the examples thus becomes
distorted. Moreover, the pragmatic decision to make all values fall between 0 and 1
may not be justified. For instance, if the difference between summer and fall is
1, it will always be bigger than, say, the difference between two normalized body
temperatures. Whether this matters or not is up to the engineer’s common sense—
assisted by his or her experience and perhaps a little experimentation.
• Why do irrelevant attributes impair the k-NN classifier’s performance? How does
the performance depend on the number of irrelevant attributes?
• Explain the basic problems pertaining to attribute scaling. Describe a simple
approach to normalization.
3.4 Performance Considerations
Having spent all this time exploring various aspects of the k-NN classifier, the
reader is bound to ask: should I really care? Sure enough, the technique is easy
to implement in a computer program, and its function is easy to grasp. But is there
a reason to believe that its classification performance is good enough?
The 1-NN Classifier Versus Ideal Bayes Let us give the question some thought.
The ultimate yardstick by which to measure k+-NN’s potential is the Bayesian
approach. We will recall that if the probabilities and pdf ’s employed by the Bayesian
formula are known with absolute accuracy, then this classifier—let us call it Ideal
Bayes—exhibits the lowest error rate theoretically achievable on the given (noisy)
data. It would be reassuring to know that the k-NN paradigm does not trail too far
behind.
The question has been subjected to rigorous mathematical analysis, and here
are the results. Figure 3.4 shows what the comparison will look like under certain
ideal circumstances such as an infinitely large training set which fills the instance
space with infinite density. The solid curve represents the two-class case where each
example is either positive or negative. We can see that if the error rate of the Ideal
Bayes is 5%, the error rate of the 1-NN classifier (vertical axis) is 10%. With the
growing amount of noise, the difference between the two approaches decreases,
only to disappear when the Ideal Bayes suffers 50% error rate—in which event, of
course, the labels of the training examples are virtually random, and any attempt at
automated classification is doomed anyway. The situation is not any better in multi-
class domains, represented in the graph by the dotted curve. Again, the Ideal Bayes
outperforms the 1-NN classifier by a comfortable margin.
3.4 Performance Considerations 53
Increasing the Number of Neighbors From the perspective of the 1-NN classifier,
the results from Fig. 3.4 are rather discouraging. On the other hand, we know that
the classifier’s performance might improve when we use the more general k-NN (for
k > 1), where some of the noisy nearest neighbors get outvoted by better-behaved
ones. Does mathematics lend some support to this expectation?
Yes it does. Under the above-mentioned ideal circumstances, the error rate has
been proven to decrease with the growing value of k, until it converges to that of
Ideal Bayes for k ! 1. At least in theory, then, the performance of the nearest-
neighbor classifier is able to reach the absolute maximum.
Practical Limitations of Theories The engineer’s world is indifferent to theoret-
ical requirements. In a realistic application, the training examples will but sparsely
populate the instance space, and increasing the number of voting neighbors can
be counterproductive. More often than not, the error rate does improve with the
growing k, but only up to a certain point from which it starts growing again—as
illustrated in Fig. 3.5 where the horizontal axis represents the values of k, and the
vertical axis represents the error rate that is measured on an independent testing set.
The explanation is simple: the more distant “nearest neighbors” may find
themselves in regions (in the instance space) that are already occupied by other
classes; as such, they only mislead the classifier. Consider the extreme: if the training
0.8
0.7
ERROR RATE OF NEAREST NEIGHBORS
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
BAYE SIAN ERROR RATE
Fig. 3.4 The theoretical error rate of the 1-NN rule compared to that of the Ideal Bayes. Two
classes: the solid curve; many classes: the dotted curve
Fig. 3.5 With the growing error rate

number of voting neighbors
(k), the error rate of the k-NN
classifier decreases until it
reaches a level from which it
starts growing again
set contains 25 training examples, then the 25-NN classifier is useless because it
simply labels any object with the class that has the highest number of representatives
in the training data.3
The Curse of Dimensionality Obviously, when classifying object x, some of its
nearest neighbors may not be similar enough to x to deserve to participate in the
vote. This often happens in domains marked by many attributes. Suppose that
the values of each attribute are confined to the unit-length interval, Œ0I 1. Using
the pythagorean theorem, it would be easy to show that the maximum Euclidean p
distance in the n-dimensional space defined by these attributes is dMAX D n. For
n D 104 (quite reasonable in such domains as, say, text categorization), this means
dMAX D 100. In view of the fact that no attribute value can exceed 1, this is perhaps
surprising. No wonder that examples then tend to be sparse—unless the training set
is really very large.
The term sometimes mentioned, in this context, is the curse of dimensionality:
as we increase the number of attributes, the number of training examples needed to
fill the instance space with adequate density grows very fast, perhaps so fast as to
render the nearest-neighbor paradigm impractical.
To Conclude Although the ideal k-NN classifier is capable of reaching the
performance of the Ideal Bayes, the engineer has to be aware of the practical
limitations of both approaches. Being able to use the Ideal Bayes is unrealistic in the
absence of the perfect knowledge of the probabilities and pdf ’s. On the other hand,
the k-NN classifier is prone to suffer from sparse data, from irrelevant attributes, and
from scaling-related problems. The concrete choice has to be based on the specific
requirements of the given application.
3
The optimal value of k (the one with the minimum error rate) is usually established experimentally.
3.5 Weighted Nearest Neighbors 55
• How does the performance of the k-NN classifier compare to that of the Ideal
Bayes? Summarize this separately for k D 1 and k > 1. What theoretical
assumptions do these two paradigms rely on?
• How will the performance of a k-NN classifier depend on the growing values of
k in theory and in a realistic application?
• What is understood by the curse of dimensionality?
3.5 Weighted Nearest Neighbors
So far, the voting mechanism was assumed to be democratic, each nearest neighbor
having the same vote. This seems to be a fair enough arrangement, but from the
perspective of classification performance, we can do better.
In Fig. 3.6, the task is to determine the class of object x, represented by a black
dot. Since three of the nearest neighbors are negative, and only two are positive,
the 5-NN classifier will label x as negative. And yet, something seems to be wrong,
here: the three negative neighbors are quite distant from x; as such, they may not
deserve to have the same say as the two positive examples in the object’s immediate
vicinity.
Weighted Nearest Neighbors Situations of this kind motivate the introduction
of weighted voting, in which the weight of each of the nearest neighbors is made
proportional to its distance from x: the closer the neighbor, the greater its impact.
Let us denote the weights as w1 ; : : : wk . The weighted k-NN classifier sums up
the weights of those neighbors that recommend the positive class (let the result be
denoted by †C ) and then sums up the weights of those neighbors that support the
negative class († ). The final verdict is based on which is higher: if †C > † ,
then x is deemed positive, otherwise it is deemed negative. Generalization to the
case with n > 2 classes is straightforward.
For the sake of illustration, suppose the positive label is found in neighbors with
weights 0:6 and 0:7, respectively, and the negative label is found in neighbors with
weights 0:1; 0:2; and 0:3. The weighted k-NN classifier will choose the positive
Fig. 3.6 The testimony of 5-NN classifier

the two “positive” neighbors
should outweigh that of the - -
three more distant “negative”
neighbors
-
++
-
+ +
class because the combined weight of the positive neighbors, †C D 0:6C0:7 D 1:3,
is greater than that of the negative neighbors, † D 0:1 C 0:2 C 0:3 D 0:6. Just
as in Fig. 3.6, the more frequent negative label is outvoted by the positive neighbors
because the latter are closer to x.
A Concrete Formula Let us introduce a simple formula to calculate the weights.
Suppose the k neighbors are ordered according to their distances, d1 ; : : : ; dk , from x
so that d1 is the smallest distance and dk is the greatest distance. The weight of the
i-th closest neighbor is calculated as follows:
(
dk di
dk d1
; dk ¤ d1
wi D (3.4)
1 dk D d1
Obviously, the weights thus obtained will range from 0 for the most distant
neighbor to 1 for the closest one. This means that the approach actually considers
only k1 neighbors (because wk D 0). Of course, this makes sense only for k > 3. If
we used k D 3, then only two neighbors would really participate, and the weighted
3-NN classifier would degenerate into the 1-NN classifier.
Table 3.4 illustrates the procedure on a simple toy domain.
Another important thing to observe is that if all nearest neighbors have the same
distance from x, then they all get the same weight, wi D 1, on account of the bottom
part of the formula in Eq. (3.4). The reader will easily verify that dk D d1 if and
only if all the k nearest neighbors have the same distance from x.
Table 3.4 Illustration of the weighted nearest neighbor rule

The task is to use the weighted 5-NN classifier to determine the class of object x. Let the
distances between x and the five nearest neighbors be d1 D 1; d2 D 3; d3 D 4; d4 D 5; d5 D 8.
Since the minimum is d1 D 1 and the maximum is d5 D 8, the individual weights are calculated
as follows:
d5 di 8 di 8 di
wi D D D
d5 d1 81 7
This gives the following values:
81 83
w1 D 7
D 1, w2 D 7
D 57 , w3 D 84
7
D 47 , w4 D 85
7
D 37 , w5 D 88
7
D 0.
If the two nearest neighbors are positive and the remaining three are negative, then x is classified
as positive because †C D 1 C 57 > † D 47 C 37 C 0.
3.6 Removing Dangerous Examples 57
• Why did we argue that each of the voting nearest neighbors should sometimes
have a different weight?
• Discuss the behavior of the formula recommended in the text for the calculation
of the individual weights.
3.6 Removing Dangerous Examples
As far as the classification procedure is concerned, the value of each individual

training example can be different. Some are typical of the classes they represent,
others less so, and yet others are outright misleading. This is why it is often a good
thing to pre-process the training set before using it: to remove examples suspected
of being less than useful.
The concrete method to pre-process the training set is guided by two essential
observations that are illustrated in Fig. 3.7. First, an example labeled with one class
Fig. 3.7 Illustration of two 2nd attribute

potentially harmful types of
examples: those surrounded
by examples of a different
− −−− − borderline region
class, and those in the + −−
“borderline region.”
+ −+ − −
+ + + −+ − −
+ +
+ + −−
+ −
−− −
1st attribute
suspected class−label noise
but surrounded by examples of another class is likely to be the result of class-

label noise. Second, examples from the borderline region between two classes are
unreliable because even small amount of noise in their attribute values can shift
their locations in the wrong directions, thus affecting classification. In both cases,
the examples better be removed.
The Technique of Tomek Links Before the culprit can be removed, however, it has
to be detected. One way to do so is by the technique of Tomek Links, named so after
the mathematician who first used them a few decades ago.4 A pair of examples, x
and y, are said to form a Tomek Link if the following three requirements are satisfied
at the same time:
1. x is the nearest neighbor of y
2. y is the nearest neighbor of x
3. the class of x is not the same as the class of y.
These conditions being characteristic of borderline examples, and also of exam-
ples surrounded by examples of anotherclass, it makes sense to remove from the
Fig. 3.8 Dotted lines connect

all Tomek links. Each y _
_
participant in a Tomek Link is
+ __ +
its partner’s nearest neighbor, _
and each of the two examples +
has a different class + _
+
+ _
+ __
+
+
x
training set all such pairs. Even this may not be enough. Sometimes, the removal of
existing Tomek Links only creates new ones so that the procedure has to be repeated.
The algorithm is summarized by the pseudocode in Table 3.5, and a few instances
of Tomek Links are shown in Fig. 3.8. Note that there are only these three; no other
pair of examples satisfies here the criteria for being called a Tomek Link.
One side-effect perhaps deserves to be mentioned: once the training set has been
cleaned, the number (k) of the voting nearest neighbors can be reduced because the
main reason for using a k > 1 is to mitigate the negative impact of noise—which
the removal of Tomek Links has reduced. It can even happen that the 1-NN classifier
will now be able to achieve the performance of a k-NN classifier that uses the entire
original training set.
A Limitation Nothing is perfect. The technique of Tomek Links does not identify
all misleading examples; and, conversely, some of the removed examples might have
been “innocent,” and thus deserved to be retained. Still, experience has shown that
the removal of Tomek Links usually does improve the overall quality of the data.
4
It is fair to mention that he used them for somewhat different purposes.
3.7 Removing Redundant Examples 59
Table 3.5 The algorithm to identify (and remove) Tomek Links

Input: the training set of N examples
1. Let i D 1 and let T be an empty set.

2. Let x be the i-th training example and let y be the nearest neighbor of x.
3. If x and y belong to the same class, go to 5.
4. If x is the nearest neighbor of y, let T D T [ fx; yg.
5. Let i D i C 1. If i N, goto 2.
6. Remove from the training set all examples that are now in T.
The engineer only has to be careful in two specific situations. (1) When the training
set is very small, and (2) when one of the classes significantly outnumbers the other.
The latter case will be discussed in Sect. 10.2.
• What motivates the attempts to “clean” the training set?
• What are Tomek Links, are how do we identify them in the training set? Why
does the procedure sometimes have to be repeated?
• How does the removal of Tomek Links affect the k-NN classifier?
3.7 Removing Redundant Examples
Some training examples do not negatively affect classification performance, and yet
we want to get rid of them. Thing is, they do not reduce the error rate, either; they
only add to computational costs.
Redundant Examples To find the nearest neighbor in a domain with 106 training
examples described by 104 attributes (and such domains are not rare), the program
relying on the Euclidean distance has to carry out 106 104 D 1010 arithmetic
operations. When the task is to classify thousands of objects at a time (which, again,
is not impossible), the number of arithmetic operations is 1010 103 D 1013 . This
is a lot.
Fortunately, training sets are often redundant in the sense that the k-NN clas-
sifier’s behavior will not change even if some of the training examples are deleted.
Sometimes, a great majority of the examples can be removed with impunity because
they add to classification costs without affecting classification performance. This is

the case of the domain shown in the upper-left corner of Fig. 3.9.
Consistent Subset In an attempt to reduce redundancy, we want to replace the
training set, T, with its consistent subset, S. In our context, S is said to be a consistent
subset of T if replacing T with S does not affect what class labels are returned by
the k-NN classifier. Such definition, however, is not very practical because we have
no idea of how the k-NN classifier will behave—using either T or S—on future
examples. Let us therefore modify the criterion: S will be regarded as a consistent
subset of T if any ex 2 T receives the same label no matter whether the k-NN
classifier has employed T fexg or S fexg.
It is in the nature of things that a realistic training set will have many consistent
subsets to choose from. Of course, the smaller the subset, the better. But a
perfectionist who insists on always finding the smallest one should be warned: this
ideal can often be achieved only at the price of enormous computational costs.
The practically minded engineer doubts that these costs are justified, and will be
content with a computationally efficient algorithm that downsizes the original set to
a “reasonable size,” unscientific though such formulation may appear to be.
Creating a Consistent Subset One such pragmatic technique is presented in
Table 3.6. The algorithm starts by choosing one random example from each class.
+ + + +
+ + + + +
+ + +++ + + + + +
+ + + + + ++ +
+ ++ + + + + ++ + + +
+ + + +
+ + + +
+
1 2
+ + + 1: original training set

+ +
+
+ 2: borderline and noisy
+ + examples removed
+ +
3: redundant examples
removed
3
Fig. 3.9 An illustration of what happens when the borderline, noisy, and redundant examples are
removed from the training set
Table 3.6 Algorithm to create a consistent training subset by the removal of (some) redundant
examples
1. Let S contain one positive and one negative example from the training set, T.
2. Using examples from S, re-classify the examples in T with the 1-NN classifier. Let M be the
set of those examples that have in this manner received the wrong class.
3. Copy to S all examples from M.
4. If the contents of S have not changed, in the previous step, then stop; otherwise go to step 1.
This initial subset, S, is then used by the 1-NN classifier to suggest the labels of all
training examples. At this stage, it is more than likely that some training examples
will thus be misclassified. These misclassified examples are added to S, and the
whole procedure is repeated using this larger set. This is then repeated again and
again until, at a certain moment, S becomes representative enough to allow the 1-
NN classifier to label all training examples correctly. This is when the search stops.
• What is the benefit of removing redundant examples from the training set?
• What do we mean by the term, “consistent subset of the training set”? Why is it
not necessary always to look for the smallest consistent subset?
• Explain the principle of the simple algorithm that creates a reasonably sized
consistent subset.
• When classifying object x, the k-NN classifier finds, in the training set, k
examples most similar to x, and then chooses the class label most common among
these examples.
• Classification behavior of the k-NN classifier depends to a great extent on how
similarities between attribute vectors are calculated. Perhaps the simplest way to
establish the similarity between vectors x and y is to use their geometric distance
obtained by the following formula:
q
dM .x; y/ D †niD1 d.xi ; yi / (3.5)
Essentially, we use d.xi ; yi / D .xi yi /2 for continuous attributes, whereas for

discrete attributes, we put d.xi ; yi / D 0 if xi D yi and d.xi ; yi / D 1 if xi ¤ yi .
• The performance of the k-NN classifier is poor if many of the attributes

describing the examples are irrelevant. Another issue is the scaling of the attribute
values. The latter problem can be mitigated by normalizing the attribute values
to unit intervals.
• Some examples are harmful in the sense that their presence in the training set
increases error rate. Others are redundant in that they only add to the computation
costs without improving classification performance. Harmful and redundant
examples should be removed.
• In some applications, each of the nearest neighbors can have the same vote. In
others, the votes are weighted based on the distance of the examples from the
classified object.
Historical Remarks The idea of the nearest-neighbor classifier was originally
proposed by Fix and Hodges [27], but the first systematic analysis was offered
by Cover and Hart [20] and Cover [19]. Exhaustive treatment was then provided
by the book by Dasarathy [21]. The weighted k-NN classifier described here was
proposed by Dudani [23]. The oldest technique to find a consistent subset of the
training set was described by Hart [35]—the one introduced in this chapter is its
minor modification. The notion of Tomek Links is due to Tomek [89].
Exercises
1. Determine the class of y D Œ1; 22 using the 1-NN classifier and the 3-NN
classifier, both using the training examples from Table 3.7. Explain why the two
classifiers differ in their classification behavior.
2. Use the examples from Table 3.7 to classify object y D Œ3; 3 with the 5-
NN classifier. Note that two nearest neighbors are positive and three nearest
neighbors are negative. Will weighted 5-NN classifier change anything? To see
what is going on, plot the locations of the examples in a graph.
Table 3.7 A simple set of x1 1 1 1 2 3 3 3 4 5

training examples for the
x2 1 2 4 3 0 2 5 4 3
exercises
class + + + +
3. Again, use the training examples from Table 3.7. (a) Are there any Tomek links?
(b) can you find a consistent subset of the training set by the removal of at least
one redundant training example?
1. Discuss the possibility of applying the k-NN classifier to the “pies” domain. Give
some thought to how many nearest neighbors to use, and what distance metric
to employ, whether to make the nearest neighbors’ vote depend on distance, and
so on.
2. Suggest other variations on the nearest-neighbor principle. Use the following
hints:
(a) Introduce alternative distance metrics. Do not forget that they have to satisfy
the axioms mentioned at the end of Sect. 3.2.
(b) Modify the voting scheme by assuming that some examples have been
created by a knowledgeable “teacher,” whereas others have been obtained
from a database without any consideration given to their representativeness.
The teacher’s examples should obviously carry more weight.
3. Design an algorithm that uses hill-climbing search to remove redundant exam-
ples. Hint: the initial state will contain the entire training set, the search operator
will remove a single training example at a time (this removal must not affect
behavior).
4. Describe an algorithm that uses hill-climbing search to remove irrelevant
attributes. Hint: withhold some training examples on which you will test 1-NN’s
classifier’s performance for different subsets of attributes.
1. Write a program whose input is the training set, a user-specified value of k, and
an object, x. The output is the class label of x.
2. Apply the program implemented in the previous assignment to some of the
benchmark domains from the UCI repository.5 Always take 40% of the examples
out and reclassify them with the 1-NN classifier that uses the remaining 60%.
3. Create a synthetic toy domain consisting of 1000 examples described by a pair
of attributes, each from the interval [0,1]. In the square defined by these attribute
values, Œ0; 1 Œ0; 1, define a geometric figure of your own choice and label all
examples inside it as positive and all other examples as negative. From this initial
5
www.ics.uci.edu/~mlearn/MLRepository.html.
noise-free data set, create 5 files, each obtained by changing p percent of the class
labels, using p 2 f5; 10; 15; 20; 25g (thus obtaining different levels of class-label
noise).
Divide each data file into two parts, the first to be reclassified by the k-NN
classifier that uses the second part. Observe how different values of k result in
different behaviors under different levels of class-label noise.
4. Implement the Tomek-link method for the removal of harmful examples. Repeat
the experiments above for the case where the k-NN classifier uses only examples
that survived this removal. Compare the results, observing how the removal
affected the classification behavior of the k-NN classifier for different values of
k and for different levels of noise.
Chapter 4
Inter-Class Boundaries: Linear and Polynomial
Classifiers
When representing the training examples with points in an n-dimensional instance

space, we may realize that positive examples tend to be clustered in regions different
from those occupied by negative examples. This observation motivates yet another
approach to classification. Instead of the probabilities and similarities employed by
the earlier paradigms, we can try to identify the decision surface that separates the
two classes. A very simple possibility is to use to this end a linear function. More
flexible are high-order polynomials which are capable of defining very complicated
inter-class boundaries. These, however, have to be handled with care.
The chapter introduces two simple mechanisms for induction of linear classifiers
from examples described by boolean attributes, and then discusses how to use them
in more general domains such as those with numeric attributes and more than just
two classes. The whole idea is then extended to polynomial classifiers.
4.1 The Essence
To begin with, let us constrain ourselves to boolean domains where each attribute is
either true or false. To make it possible for these attributes to participate in algebraic
functions, we will represent them by integers: true by 1, and false by 0.
The Linear Classifier In Fig. 4.1, one example is labeled as positive and the
remaining three as negative. In this particular case, the two classes can be separated
by the linear function defined by the following equation:
1:2 C 0:5x1 C x2 D 0 (4.1)
In the expression on the left-hand side, x1 and x2 represent attributes. If we

substitute for x1 and x2 the concrete values of a given example, the expression,
1:2C0:5x1 Cx2 , will become either positive or negative. This sign then determines

DOI 10.1007/978-3-319-63913-0_4
66 4 Inter-Class Boundaries: Linear and Polynomial Classifiers
x2
−1.2 + 0.5 x1 + x2 = 0
1.2
− + x1 x2 − 1.2+0 .5 x1 + x2 Class
1
1 1 0.3 pos
1 0 −0.7 neg
0 1 −0.2 neg
0 0 −1.2 neg
0 − −
0 1 x1
Fig. 4.1 A linear classifier in a domain with two classes and two boolean attributes (using 1 instead
of true and 0 instead of false)
the example’s class. The table on the right shows how the four examples from the
left have thus been classified.
Equation (4.1) is not the only one capable of doing the job. Other expressions,
say, 1:5 C x1 C x2 , will label the four examples in exactly the same way. As a
matter of fact, the same can be accomplished by infinitely many different classifiers
of the following generic form:
w0 C w1 x1 C w2 x2 D 0
This is easy to generalize to domains with n attributes:
w0 C w1 x1 C : : : C wn xn D 0 (4.2)
If n D 2, Eq. (4.2) defines a line; if n D 3, a plane; and if n > 3, a hyperplane. By

the way, if we artificially introduce a “zeroth” attribute, x0 , whose value is always
fixed at x0 D 1, the equation can be re-written in the following, more compact,
form:
X
n
wi xi D 0 (4.3)
iD0
Some Practical Issues When writing a computer program implementing the

classifier, the engineer must not forget to decide how to label the rare example that
finds itself exactly on the hyperplane—which happens when the expression equals 0.
Common practice either chooses the class randomly or gives preference to the one
that has more representatives in the training set.
Also, we must not forget that no linear classifier can ever separate the positive
and the negative examples if the two classes are not linearly separable. Thus if we
change in Fig. 4.1 the class label of x D .x1 ; x2 / D .0; 0/ from minus to plus,
4.1 The Essence 67
no straight line will ever succeed. Let us defer further discussion of this issue till
Sect. 4.5. For the time being, we will consider only domains where the classes are
linearly separable.
The Parameters The classifier’s behavior is determined by the coefficients, wi .
These are usually called weights. The task for machine learning is to find their
appropriate values.
Not all the weights play the same role. Geometrically speaking, the coefficients
w1 ; : : : wn define the angle of the hyperplane with respect to the system of
coordinates; and the last, w0 , called bias, determines the offset, the hyperplane’s
distance from the origin of the system of coordinates.
The Bias and the Threshold In the case depicted in Fig. 4.1, the bias was w0 D
1:2. A higher value would “shift” the classifier further from the origin, Œ0; 0,
whereas w0 D 0 would make the classifier intersect the origin. Our intuitive grasp
of the role played by bias in the classifier’s behavior will improve if we re-write
Eq. (4.2) as follows:
w1 x1 C : : : wn xn D (4.4)
The term on the right-hand side, D w0 , is the threshold that the weighted
sum has to exceed if the example is to be deemed positive. Note that the threshold
equals the negatively taken bias.
Simple Logical Functions Let us simplify our considerations by the (somewhat
extreme) requirement that all attributes should have the same weight, wi D 1. Even
under this constraint, careful choice of the threshold will implement some useful
functions. For instance, the reader will easily verify that the following classifier
returns the positive class if and only if every single attribute has xi D 1, a situation
known as logical AND.
x1 C : : : C xn D n 0:5 (4.5)
By contrast, the next classifier returns the positive class if at least one attribute is
xi D 1, a situation known as logical OR.
x1 C : : : C xn D 0:5 (4.6)
Finally, the classifier below returns the positive class if at least k attributes (out
of the total of n attributes) are xi D 1. This represents the so-called k-of-n function,
of which AND and OR are special cases: AND is n-of-n, whereas OR is 1-of-n.
x1 C : : : C xn D k 0:5 (4.7)
Weights Now that we understand the impact of bias, let us abandon the restriction
that all weights be equal, and take a look at the consequences of their concrete
values. Consider the linear classifier defined by the following function:
2 C 3x1 2x2 C 0:1x4 0:5x6 D 0 (4.8)
The first thing to notice in the expression on the left side is the absence of
attributes x3 and x5 . These are rendered irrelevant by their zero weights, w3 D w5 D
0. As for the other attributes, their impacts depend on their weights’ absolute values
as well as on the signs: if wi > 0, then xi D 1 increases the chances of the above
expression being positive; and if wi < 0, then xi D 1 increases the chances of the
expression being negative. Note that, in the case of the classifier defined by Eq. (4.8),
x1 supports the positive class more strongly than x4 because w1 > w4 . Likewise, the
influence of x2 is stronger than that of x6 —only in the opposite direction: reducing
the value of the overall sum, this weight makes it more likely that an example with
x2 D 1 will be deemed negative. Finally, the very small value of w4 renders attribute
x4 almost irrelevant.
As another example, consider the classifier defined by the following function:
2x1 C x2 C x3 D 1:5 (4.9)
Here the threshold 1:5 is exceeded either by the sole presence of x1 D 1 (because
then 2x1 D 2 1 > 1:5) or by the combined contributions of x2 D 1 and x3 D 1.
This means that x1 will “outvote” the other two attributes even when x2 and x3 join
their forces in the support of the negative class.
Low Computational Costs Note the relatively low computational costs of this
approach. Whereas the 1-NN classifier had to evaluate many geometric distances,
and then search for the smallest among them, the linear classifier only has to
determine the sign of a relatively simple expression.
• Write the general expression defining the linear classifier in a domain with four
boolean attributes. Why do we prefer to represent the true and false values by 1
and 0, respectively? How does the classifier determine an example’s class?
• How can a linear classifier implement functions AND, OR, and k-of-n?
• Explain and discuss the behavior of the linear classifier defined by the expression,
2:5 C x2 C 2x3 D 0. What do the weights tell us about the role of the individual
attributes?
• Compare the computational costs of the linear classifier with those of the nearest-
neighbor classifier.
4.2 The Additive Rule: Perceptron Learning 69
Fig. 4.2 The linear classifier h( Σ w i x i )

outputs h.x/ D 1 when
P n
wi xi > 0 and h.x/ D 0
iD0 P
n
when iD0 wi xi 0,
signaling the example is 1
positive or negative,
respectively
0 Σ wi xi
4.2 The Additive Rule: Perceptron Learning
Having developed some understanding of how the linear classifier works, we are
ready to take a closer look at how to induce the tool from training data.
The Learning Task We will assume that each training example, x, is described by
n binary attributes whose values are either xi D 1 or xi D 0. A positive example
is labeled with c.x/ D 1, and a negative with c.x/ D 0. To make sure we never
confuse an example’s real class with the one returned by the classifier, we will
denote
P the latter by h.x/ where h indicates that this is the classifier’s hypothesis.
If niD0 wi xi > 0, the classifier “hypothesizes”
Pn that the example is positive, and
therefore returns h.x/ D 1. Conversely, if iD0 wi xi 0, the classifier returns
h.x/ D 0. Figure 4.2 serves as a reminder that the classifier labels x as positive only
if the cumulative evidence in favor of this class exceeds 0.
Finally, we will suppose that examples with c.x/ D 1 are linearly separable from
those with c.x/ D 0. This means that there exists a linear classifier that will label
correctly all training examples, h.x/ D c.x/. The task for machine learning is to
find weights, wi , that will make this happen.
Learning from Mistakes Here is the essence of the most common approach to
the induction of linear classifiers. Suppose we already have an interim (even if
imperfect) version of the classifier. When presented with a training example, x, the
classifier returns its label, h.x/. If this differs from the true class, h.x/ ¤ c.x/, it is
because the weights are less than perfect; they thus have to be modified in a way
likely to correct this error.
Here is how to go about the weight modification.
P Let the true class be c.x/ D 1.
In this event, h.x/ D 0 will only happen if niD0 wi xi < 0, an indication P that the
weights are too small. If we increase the weights, then the sum, niD0 wi xi , may
exceed zero, making the returned label positive, and thus correct. Note that it is
enough to increase only the weights of attributes with xi D 1; when xi D 0, then
the value of wi does not matter because anything multiplied by zero is still zero:
0 wi D 0.
Similarly, if c.x/ D 0 and h.x/ D 1, then the weights of all attributes such that
xP i D 1 should be decreased so as to give the sum the chance to drop below zero,
n
iD0 wi xi < 0.
The Weight-Adjusting Formula In summary, the presentation of a training

example, x, can result in three different situations. The technique based on “learning
from mistakes” responds to them according to the following table:
Situation Action
c.x/ D 1 while h.x/ D 0 Increase wi for each attribute with xi D 1
c.x/ D 0 while h.x/ D 1 Decrease wi for each attribute with xi D 1
c.x/ D h.x/ Do nothing
Interestingly, all these actions are carried out by a single formula:
wi D wi C Œc.x/ h.x/ xi (4.10)
Let us take a look at the basic aspects of this weight-adjusting formula.

1. Correct action. If c.x/ D h.x/, the term in the brackets is Œc.x/ h.x/ D 0,
which leaves wi unchanged. If c.x/ D 1 and h.x/ D 0, the term in the brackets is
1, and the weights are thus increased. And if c.x/ D 0 and h.x/ D 1, the term in
the brackets is negative, and the weights are reduced.
2. Affecting only relevant weights. If xi D 0, the term to be added to the i-th weight,
Œc.x/ h.x/ 0, is zero. This means that the formula will affect wi only when
xi D 1.
3. Amount of change. This is controlled by the learning rate, , whose user-set value
is chosen from the interval 2 .0; 1.
Note that the modification of the weights is additive in the sense that a term is
added to the previous value of the weight. In Sect. 4.3, we will discuss the other
possibility: a multiplicative formula.
The Perceptron Learning Algorithm Equation (4.10) forms the core of the
Perceptron Learning Algorithm.1 The procedure is summarized by the pseudocode
in Table 4.1. The principle is simple. Once the weights have been initialized to
small random values, the training examples are presented, one at a time. After each
example presentation, every weight of the classifier is subjected to Eq. (4.10). The
last training example signals that one training epoch has been completed. Unless
the classifier now labels correctly the entire training set, the learner returns to the
first example, thus beginning the second epoch, then the third, and so on. Typically,
several epochs are needed to reach the goal.
A Numeric Example Table 4.2 illustrates the procedure on a toy domain where
the three training examples, e1 , e2 , and e3 , are described by two binary attributes.
After the presentation of e1 , the weights (originally random) are reduced on account
1
Its author, M. Rosenblatt, originally employed this learning technique in a device he called a
Perceptron.
4.2 The Additive Rule: Perceptron Learning 71
Table 4.1 The perceptron learning algorithm

Assumption: the two classes, c.x/ D 1 and c.x/ D 0, are linearly separable.
1. Initialize all weights, wi , to small random numbers.
Choose an appropriate learning rate, 2 .0; 1.
2. For each training example, x D .x1 ; : : : ; xn /, whose class is c.x/:
Pn
(i) Let h.x/ D 1 if iD0 wi xi > 0, and h.x/ D 0 otherwise.
(ii) Update each weight using the formula, wi D wi C Œc.x/ h.x/ xi
3. If c.x/ D h.x/ for all training examples, stop; otherwise, return to step 2.
Table 4.2 Illustration of perceptron learning

Let the learning rate be D 0:5, and let the (randomly generated) initial weights be w0 D
0:1; w1 D 0:3, and w3 D 0:4. Set x0 D 1.
Task: Using the following training set, the perceptron learning algorithm is to learn how to
separate the negative examples, e1 and e3 , from the positive example, e2 .
Example x1 x2 c.x/
e1 1 0 0
e2 1 1 1
e3 0 0 0
Pn
The linear classifier’s hypothesis about x’s class is h.x/ D 1 if iD0 wi xi > 0 and h.x/ D 0
otherwise. After each example presentation, all weights are subjected to the same formula:
wi D wi C 0:5 Œc.x/ h.x/ xi .
The table below shows, step by step, what happens to the weights in the course of learning.
x1 x2 w0 w1 w2 h.x/ c.x/ c.x/ h.x/

Random classifier 0.1 0.3 0.4
Example e1 1 0 1 0 1
New classifier 0.4 0.2 0.4
New classifier 0.1 0.3 0.9
Final classifier 0.4 0.3 0.9
The final version of the classifier, 0:4 C 0:3x1 C 0:9x2 D 0, no longer misclassifies any
training example. The training has thus been completed in a single epoch.
of h.e1 / D 1 and c.e1 / D 0; however, this happens only to w0 and w1 because

x2 D 0. In response to e2 , all weights of the classifier’s new version are increased
because h.e2 / D 0 and c.e2 / D 1, and all attributes have xi D 1. And after e3 ,
the fact that h.e3 / D 1 and c.e1 / D 0 results in the reduction of w0 , but not of the
other weights because x1 D x2 D 0. From now on, the classifier correctly labels all
training examples, and the process can thus be terminated.
Initial Weights and the Number of Attributes The training converged to the
separating line in a single epoch, but this was only thanks to a few lucky choices
that may have played quite a critical role. Let us discuss them briefly.
First of them is the set of (random) initial weights. Different initialization may
result in a different number of epochs. Most of the time, the classifier’s initial version
will be almost useless, and a lot of training is needed before the process converges
to something useful. Sometimes, however, the first version may be fairly good, and
a single epoch will do the trick. And at the extreme case, there exists the possibility,
however remote, that the random-number generator will create a classifier that labels
all training examples without a single error, and no training is thus needed.
Another factor is the length of the attribute vector. As a rule of thumb, the number
of the necessary epochs tends to grow linearly in the number of attributes (assuming
the same learning rate, , is used). For instance, the number of epochs needed in a
domain with 3 n attributes is likely to be about three times the number of epochs
that would be needed in a domain with n attributes.
Learning Rate A critical role is played by the learning rate, . Returning to the
example from Table 4.2, the reader will note the rapid weight changes. Thus w0
“jumped” from 0:1 to 0:4 after e1 , then back to 0.1 after e2 , only to return to 0:4
after e3 . Similar changes were experienced by w1 and w2 . Figure 4.3 visualizes the
phenomenon. The reader will easily verify that the four lines represent the four
successive versions of the classifier. Note how dramatic, for instance, is the change
from classifier 1 to classifier 2, and then from classifier 2 to classifier 3.
Fig. 4.3 The four classifiers x2 2

from Table 4.2. The classifier − +
defined by the initial weights
is denoted by 1; numbers 2
and 3 represent the two
intermediate stages; and 4,
the final solution. The arrows
indicate the half-space of
positive examples
− x1
− 4
3
1
4.3 The Multiplicative Rule: WINNOW 73
This remarkable sensitivity is explained by the high learning rate, D 0:5. A

smaller value, such as D 0:1, would moderate the changes, thus “smoothing out”
the learning. But if we overdo it by choosing, say, D 0:001, the training process
will become way too slow, and a great many epochs will have to pass before all
training examples are correctly classified.
If the Solution Exists, It Will Be Found Whatever the initial weights, whatever
the number of attributes, and whatever the learning rate, one thing is always
guaranteed. If the positive and negative classes are linearly separable, the perceptron
learning algorithm is guaranteed to find one version of the class-separating hyper-
plane in a finite number of steps.
• Under what circumstances is perceptron learning guaranteed to find a classifier
that perfectly labels all training examples?
• When does the algorithm reduce the classifier’s weights, when does it increase
them, and when does it leave them unchanged? Why does it modify wi only if
the corresponding attribute’s value is xi D 1?
• What circumstances influence the number of epochs needed by the algorithm to
converge to a solution?
4.3 The Multiplicative Rule: WINNOW
Perceptron learning responds to the classifier’s error by applying the additive rule
that added to the weights a positive or negative term. An obvious alternative is
the multiplicative rule where the weights are multiplied instead of being added to.
Such approach has been adopted by WINNOW, an algorithm summarized by the
pseudocode in Table 4.3.
The Principle and the Formula The general scenario is the same as in the
previous section. A training example, x, is presented, and the classifier returns its
hypothesis about the example’s label, h.x/. The learner compares this hypothesis
with the known label, c.x/. If the two differ, c.x/ ¤ h.x/, the weights of the
attributes with xi D 1 are modified in the following manner (where ˛ > 1 is a
user-set parameter):
Situation Action
c.x/ D 1 while h.x/ D 0 wi D ˛wi
c.x/ D 0 while h.x/ D 1 wi D wi =˛
c.x/ D h.x/ Do nothing
Table 4.3 The WINNOW algorithm

Assumption: the two classes, c.x/ D 1 and c.x/ D 0, are linearly separable.
1. Initialize the classifier’s weights to wi D 1.
2. Set the threshold to D n 0:1 (n being the number of attributes) and choose an
appropriate value for parameter ˛ > 1 (usually ˛ D 2).
3. Present a training example, x, whose class is c.x/. The classifier returns h.x/.
4. If c.x/ ¤ h.x/, update the weights of each attribute whose value is xi D 1:
if c.x/ D 1 and h.x/ D 0, then wi D ˛wi
if c.x/ D 0 and h.x/ D 1, then wi D wi =˛
5. If c.x/ D h.x/ for all training examples, stop; otherwise, return to step 3.
The reader is encouraged to verify that all these three actions can be carried out
by the same formula:
wi D wi ˛ c.x/h.x/ (4.11)
A Numeric Example Table 4.4 illustrates the principle using a simple toy domain.
The training set consists of all possible examples that can be described by three
binary attributes. Those with x2 D x3 D 1 are labeled as positive and all others as
negative, regardless of the value of the (irrelevant) attribute x1 .
In perceptron learning, the weights were initialized to small random values. In
the case of WINNOW, however, they are all initially set to 1. As for the threshold,
D n 0:1 is used, slightly less than the number of attributes. In the toy domain
from Table 4.4, this means D 3 0:1 D 2:9 because WINNOW of course has no
a priori knowledge of one of the attributes being irrelevant.
When the first four examples are presented, the classifier’s initial version labels
them all correctly. The first mistake is made in the case of e5 : for this positive
example, the classifier incorrectly returns the negative label. The learner therefore
increases the weights of attributes with xi D 1 (that is, w2 and w3 ). This new
classifier then classifies correctly all the remaining examples, e6 through e8 . In the
second epoch, the classifier errs on e2 , causing a false positive. In response to this
error, the algorithm reduces weights w1 and w2 (but not w3 because x3 D 0). After
this last weight modification, the classifier labels correctly the entire training set.2
2
Similarly as in the case of perceptron learning, we could have considered the 0-th attribute, x0 D
1, whose initial weight is w0 D 1.
4.3 The Multiplicative Rule: WINNOW 75
Table 4.4 Illustration of the WINNOW’s behavior

Task. Using the training examples from the table on the left (below), induce the linear
classifier. Let ˛ D 2, and let D 2:9.
Note that the training is here accomplished in two learning steps: presentation of e5 (false
negative), and of e2 (false positive). After these two weight modifications, the resulting
classifier correctly classifies all training examples.
x1 x2 x3 c(x)
e1 1 1 1 1
x1 x2 x3 w1 w2 w3 h.x/ c.x/
e2 1 1 0 0
Init. class. 1 1 1
e3 1 0 1 0
e4 1 0 0 0
New weights 1 2 2
e5 0 1 1 1
e6 0 1 0 0
New weights 0.5 1 2
e7 0 0 1 0
e8 0 0 0 0
Note that the weight of the irrelevant attribute, x1 , is now smaller than the weights
of the relevant attributes. Indeed, the ability to penalize irrelevant attributes by
significantly reducing their weights, thus “winnowing them out,” is one of the main
advantages of this technique.
The “Alpha” Parameter Parameter ˛ controls the learner’s sensitivity to errors
in a manner reminiscent of the learning rate in perceptron learning. The main
difference is the requirement that ˛ > 1. This guarantees an increase in weight
wi in the case of a false negative, and a decrease in wi in the case of a false positive.
The parameter’s concrete value is not completely arbitrary. If it exceeds 1 by just a
little (say, if ˛ D 1:1), then the weight-updates will be very small, resulting in slow
convergence. Increasing ˛’s value accelerates convergence, but risks overshooting
the solution. The ideal value is best established experimentally; good results are
often achieved with ˛ D 2.
No Negative Weights? Let us point out one fundamental difference between
WINNOW and perceptron learning. Since the (originally positive) weights are
always multiplied by ˛ or 1=˛, none of them can ever drop to zero, let alone
become negative. This means that, unless appropriate measures have been taken,
a whole class of linear classifiers will thus be eliminated: those with negative or
zero coefficients.
The shortcoming is removed if we represent each of the original attributes by a
pair of “new” attributes: one copying the original attribute’s value, the other having
the opposite value. In a domain that originally had n attributes, the total number
of attributes will then be 2n, the value of the .n C i/th attribute, xnCi , being the
opposite of xi . For instance, suppose that an example is described by the following
three attribute values:
x1 D 1; x2 D 0; x3 D 1
In the new representation, the same example will be described by six attributes:
x1 D 1; x2 D 0; x3 D 1; x4 D 0; x5 D 1; x6 D 0;
For these, WINNOW will have to find six weights, w1 ; : : : ; w6 , or perhaps seven,
if w0 is used.
Comparing It with Perceptron In comparison with perceptron learning, WIN-
NOW appears to converge faster in domains with irrelevant attributes whose weights
are quickly reduced to small values. However, neither WINNOW nor perceptron
learning is able to recognize (and eliminate) redundant attributes. In the event of
two attributes always having the same value, xi D xj , the learning process will
converge to the same weight for both, making them look equally important even
though it is clear that only one of them is strictly needed.
• What formula is used by the weight-updating mechanism in WINNOW? Why is
the formula called multiplicative?
• What is the shortcoming of multiplying or dividing the weights by ˛ > 1? How
is the situation remedied?
• Summarize the differences between WINNOW and perceptron learning.
4.4 Domains with More Than Two Classes
Having only two sides, a hyperplane may separate the positive examples from
the negative—and that’s it; when it comes to multi-class domains, the tool seems
helpless. Or is it?
Groups of Binary Classifiers What is beyond the powers of an individual can
be solved by a team. One approach that is sometimes employed, in this context,
is illustrated in Fig. 4.4. The “team” consists of four binary classifiers, each
specializing on one of the four classes, C1 through C4 . Ideally, the presentation
of an example from Ci results in the i-th classifier returning hi .x/ D 1, and all the
other classifiers returning hj .x/ D 0, assuming, again, that each class can be linearly
separated from the other classes.
Modifying the Training Data To exhibit this behavior, however, the individual
classifiers need to be properly trained. Here, one can rely on the algorithms that
have been described in the previous sections. The only additional “trick” is that the
engineer needs to modify the training data accordingly.
4.4 Domains with More Than Two Classes 77
Fig. 4.4 Converting a 4-class C1 C2 C3 C4

problem into four 2-class
problems
attribute vector
Table 4.5 A 4-class training set, T, converted to 4 binary training sets, T1 : : : T4
T T1 T2 T3 T4
e1 C2 e1 0 e1 1 e1 0 e1 0
e2 C1 e2 1 e2 0 e2 0 e2 0
e3 C3 e3 0 e3 0 e3 1 e3 0
e4 C4 e4 0 e4 0 e4 0 e4 1
e5 C2 e5 0 e5 1 e5 0 e5 0
e6 C4 e6 0 e6 0 e6 0 e6 1
Table 4.5 illustrates the principle. On the left is the original training set, T, where
each example is labeled with one of the four classes. On the right are four “derived”
sets, T1 through T4 , each consisting of the same six examples which, however, have
been re-labeled so that an example that in the original set, T, represents class Ci is
labeled with c.x/ D 1 in Ti and with c.x/ D 0 in all other sets.
The Need for a Master Classifier The training sets, Ti , are presented to a learner
which induces from each of them a linear classifier dedicated to the corresponding
class. This is not the end of the story, though. The training examples may poorly
represent the classes, they may be corrupted by noise, and even the requirement
of linear separability may not be satisfied. As a result of these complications, the
induced classifiers may “overlap” each other in the sense that two or more of them
will respond to the same example, x, with hi .x/ D 1, leaving the wrong impression
that x belongs to more than one class. This is why a “master classifier” is needed,
its task being to choose from the returned classes the one most likely to be correct.
This is not difficult. The reader will recall that a linear classifier labels x as
positive if the weighted sum of x’s attribute values exceeds zero: †niD0 wi xi > 0. This
sum (usually different in each of the classifiers that have returned hi .x/ D 1) can
be interpreted as the amount of evidence in support of the corresponding class. The
master classifier then simply gives preference to the class whose binary classifier
delivered the highest †niD0 wi xi .
A Numeric Example The principle is illustrated in Table 4.6. Here, each of the
rows represents a different class (with the total of four classes). Of course, each
Table 4.6 Illustration of the master classifier’s behavior: choosing the example’s class from
several candidates
Suppose we have four binary classifiers (the i-th classifier used for the i-th class) defined
by the weights listed in the table below. How shall the master classifier label example x D
.x1 ; x2 ; x3 ; x4 / D .1; 1; 1; 0/?
Classifier
Class w0 w1 w2 w3 w4 †niD0 wi xi h.x/
C1 1.5 1 0.5 1 5 1 0
C2 0.5 1.5 1 3 1 4 1
C3 1 2 4 1 0.5 2 1
C4 2 1 1 3 1 3 0
The rightmost column tells us that two classifiers, C2 and C3 , return h.x/ D 1. From these, C2
is supported by the higher value of †niD0 wi xi . Therefore, the master classifier labels x with C2 .
classifier has a different set of weights, each weight represented by one column in
the table. When an example is presented, its attribute-values are in each classifier
multiplied by the corresponding weights. We observe that in the case of two
classifiers, C2 and C3 , the weighted sums are positive, †niD0 wi xi > 0, which might
mean that both classifiers return h.x/ D 1. Since each example is supposed to be
labeled with one and only one class, we need a master classifier to make a decision.
In this particular case, the master classifier gives preference to C2 because this
classifier’s weighted sum is greater than that of C3 .
A Practical Limitation A little disclaimer is in place here. This method of
employing linear classifiers in multi-class domains is reliable only if the number of
classes is moderate, say, 3–5. In domains with many classes, the “derived” training
sets, Ti , will be imbalanced in the sense that most examples will have c.x/ D 0 and
only a few c.x/ D 1. As we will learn in Sect. 10.2, imbalanced training sets tend to
cause difficulties in noisy domains unless appropriate measures have been taken.
• When trying to use N linear classifiers in an N-class domain, how will you create
the training sets, Ti , for the induction of the individual binary classifiers?
• How can an example’s class be decided upon in a situation where two or more
binary classifiers return h.x/ D 1?
4.5 Polynomial Classifiers 79
4.5 Polynomial Classifiers
It is now time to abandon the requirement that the positive examples be linearly
separable from the negative; because fairly often, they are not. Not only can the
linear separability be destroyed by noise. The very shape of the region occupied
by one of the classes can make it impossible to separate it by a linear decision
surface. Thus in the training set shown in Fig. 4.5, no linear classifier ever succeeds
in separating the two classes, a feat that can only be accomplished by a non-linear
curve such as the parabola shown in the picture.
Non-linear Classifiers The point having been made, we have to ask how to induce
the more sophisticated non-linear classifiers. There is no doubt that they exist. For
instance, math teaches us that any n-dimensional function can be approximated to
arbitrary precision with a polynomial of a sufficiently high order. Let us therefore
take a look at how to use—and induce—the polynomials for our classification
purposes.
Polynomials of the Second Order The good news is that the coefficients of
polynomials can be induced by the same techniques that we have used for linear
classifiers. Let us explain how.
For the sake of clarity, we will begin by constraining ourselves to simple domains
with only two boolean attributes, x1 and x2 . The second-order polynomial function
is then defined as follows:
w0 C w1 x1 C w2 x2 C w3 x12 C w4 x1 x2 C w5 x22 D 0 (4.12)
The expression on the left is a sum of terms that all have one thing in common:
a weight, wi , that multiplies a product x1k x2l . In the first term, we have k C l D 0,
because w0 x10 x20 D w0 ; next come the terms with k C l D 1, concretely, w1 x11 x20 D
w1 x1 and w2 x10 x21 D w1 x2 ; and the sequence ends with the three terms that have
k C l D 2: specifically, w3 x12 , w4 x11 x21 , and w5 x22 . The thing to remember is that the
expansion of the second-order polynomial stops when the sum of the exponents
reaches 2.
Fig. 4.5 In some domains, x2

no linear classifier can − +
separate the positive
examples from the negative.
Only a non-linear classifier
can do so
+ −
x1
Of course, some of the weights can be wi D 0, rendering the corresponding terms

“invisible” such as in 7 C 2x1 x2 C 3x22 where the coefficients of x1 ; x2 ; and x12 are
zero.
Polynomials of the r-th Order More generally, the r-th order polynomial (still in
a two-dimensional domain) will consist of a sum of weighted terms, wi x1k x2l , such
that k C l D j, where j D 0; 1; : : : r.
The reader will easily make the next step and write down the general formula
that defines the r-th order polynomial for domains with more than two attributes. A
hint: the sum of the exponents in any single term never exceeds r.
Converting Them to Linear Classifiers Whatever the polynomial’s order, and
whatever the number of attributes, the task for machine learning is to find weights
that make it possible to separate the positive examples from the negative. The
seemingly unfortunate circumstance that the terms are non-linear (the sum of the
exponents sometimes exceeds 1) is easily removed by multipliers, devices that
return the logical conjunction of inputs: 1 if all inputs are 1; and 0 if at least one
input is 0. With their help, we can replace each product of attributes with a new
attribute, zi , and thus re-write Eq. (4.12) in the following way:
w0 C w1 z1 C w2 z2 C w3 z3 C w4 z4 C w5 z5 D 0 (4.13)
This means, for instance, that z3 D x12 and z4 D x1 x2 . Note that this “trick” has
transformed the originally non-linear problem with two attributes, x1 and x2 , into a
linear problem with five newly created attributes, z1 through z5 .
Figure 4.6 illustrates the situation where a second-order polynomial is used in a
domain with three attributes.
Since the values of zi in each example are known, the weights can be obtained
without any difficulties using perceptron learning or WINNOW. Of course, we must
not forget that these techniques will find the solution only if the polynomial of the
chosen order is indeed capable of separating the two classes.
• When do we need non-linear classifiers? Specifically, what speaks in favor of
polynomial classifiers?
• Write down the mathematical expression that defines a polynomial classifier.
What “trick” allows us to use here the same learning techniques that were used
in the case of linear classifiers?
4.6 Specific Aspects of Polynomial Classifiers 81
The second-order polynomial function over three attributes is defined by the follow-
ing function:
0 = w0 + w1 x1 + w2 x2 + w3 x3 + w4 x21 + w5 x1 x2 + w6 x1 x3
+w7 x22 + w8 x2 x3 + w9 x23
Using the multipliers, we obtain the following:
0 = w0 + w1 z1 + w2 z2 + w3 z3 + w4 z4 + w5 z5 + w6 z6 + w7 z7 + w8 z8 + w9 z9
Below is the schema of the whole “device” with multipliers. Before reaching the
linear classifier, each signal zi is multiplied by the corresponding weight, wi .
z1
z2 x1
z3 x2
x3
z4 x1
linear classifier
x1
z x
1
5
x2
z6 x1
h
x3
z7 x2
x2
z8 x2
x3
z9 x3
x3
Fig. 4.6 A polynomial classifier can be converted into a linear classifier with the help of
multipliers that pre-process the data
4.6 Specific Aspects of Polynomial Classifiers
To be able to use a machine-learning technique with success, the engineer must

understand not only its strengths, but also its limitations, shortcomings, and pitfalls.
In the case of polynomial classifiers, there are a few that deserve our attention. Let
us briefly discuss them.
Fig. 4.7 The two classes are x2 x2

linearly separable, but noise
has caused one negative
example to be mislabeled as
positive. The high-order
polynomial on the right
overfits the data, ignoring the
possibility of noise
x1 x1
Overfitting Polynomial classifiers tend to overfit noisy training data. Since the
problem of overfitting is encountered also in other machine-learning paradigms, we
have to discuss its essence in some detail. For the sake of clarity, we will abandon
the requirement that all attributes should be boolean; instead, we will rely on two-
dimensional continuous domains that are easy to visualize.
The eight training examples in Fig. 4.7 fall into two groups. In one of them,
all examples are positive; in the other, all save one are negative. Two attempts at
separating the two classes are made. The one on the left is content with a linear
classifier, shrugging off the minor inconvenience that one training example remains
misclassified. The one on the right resorts to a polynomial classifier in an attempt to
avoid any error being make on the training set.
An Inevitable Trade-Off Which of the two is to be preferred? This is not an easy
question. On the one hand, the two classes may be linearly separable, and the only
cause for one positive example to be found in the negative region is class-label noise.
If this is the case, the single error made by the linear classifier on the training set is
inconsequential, whereas the polynomial, cutting deep into the negative area, will
misclassify examples that find themselves on the wrong side of the curve. On the
other hand, there is also the chance that the outlier does represent some legitimate,
even if rare, aspect of the positive class. In this event, using the polynomial will be
justified. On the whole, however, the assumption that the single outlier is nothing
but noise is more likely to be correct than the “special-aspect” alternative.
A realistic training set will contain not one, but quite a few, perhaps many
examples that seem to find themselves in the wrong part of the instance space. And
the inter-class boundary that our classifier attempts to model may indeed be curved,
though how much curved is anybody’s guess. The engineer may lay aside the linear
classifier as too crude an approximation, and opt instead for the greater flexibility
offered by the polynomials. This said, a very-high-order polynomial will avoid any
error even in a very noisy training set—and then fail miserably on future data. The
ideal solution is often somewhere between the extremes of linear classifiers and
high-order polynomials. The best choice can be determined experimentally.
The Number of Weights The total number of the weights to be trained depends
on the length of the attribute vector, and on the order of the polynomial. A simple
analysis reveals that, in the case of n attributes and the r-th order polynomial, the
number is determined by the following combinatorial expression:
4.6 Specific Aspects of Polynomial Classifiers 83

nCr
NW D (4.14)
r
Of course, NW will be impractically high for large values of n. For instance,

even for the relatively modest case of n D 100 attributes and a polynomial’s order
r D 3, the number of weights to be trained is NW D 176;851 (103 choose 3). The
computational costs thus incurred are not insurmountable for today’s computers.
What is worse is the danger of overfitting noisy training data; the polynomial is
simply too flexible. The next paragraphs will tell us how much flexible.
Capacity The trivial domain in Fig. 4.1 consisted of four examples. Given that each
of them can be labeled as either positive or negative, we have 24 D 16 different
ways of assigning labels to this training set. Of these sixteen, only two represent a
situation where the two classes cannot be linearly separated—in this domain, linear
inseparability is a rare event. But how typical is this situation in the more general
case of m examples described by n attributes’? What are the chances that a random
labeling of the examples will result in linearly separable classes?
Mathematics has found a simple guideline to be used in domains where n is
“reasonably high” (say, ten or more attributes): if the number of examples, m, is
less than twice the number of attributes (m < 2n), the probability that a random
distribution of the two labels will result in linear separability is close to 100%.
Conversely, this probability is almost zero when m > 2n. In this sense, “the capacity
of a linear classifier is twice the number of attributes.”
This result applies also to polynomial classifiers. The role of attributes is here
played by the terms, zi , obtained by the multipliers. Their number, NW , is obtained
by Eq. (4.14). We have seen that NW can be quite high—and this makes the capacity
high, too. In the case of n D 100 and r D 3, the number of weights is 176;851.
This means that the third-order polynomial can separate the two classes (regardless
of noise) as long as the number of examples is less than 353;702.
• Elaborate on the term, “overfitting” and explain why this phenomenon (and its
consequences) is in polynomial classifiers difficult to avoid.
• What is the upper bound on the number of weights to be trained in a polynomial
of the r-th order in a domain that has n attributes?
• What is the capacity of the linear or polynomial classifier? What does capacity
tell us about linear separability?
4.7 Numerical Domains and Support Vector Machines
Now that we have realized that polynomials do not call for new machine-learning
algorithms, we can return to linear classifiers, a topic we have not yet exhausted.
Time has come to abandon our self-imposed restriction to boolean attributes, and to
start considering also the possibility of the attributes being continuous. Can we then
still rely on the two training algorithms described above?
Perceptron in Numeric Domains In the case of perceptron learning, the answer is
easy: yes, the same weight-modification formula can be used. Practical experience
shows, however, that it is good to make all attribute values fall into the unit interval,
xi 2 Œ0; 1.
Let us repeat here, for the reader’s convenience, the weight-adjusting formula:
wi D wi C Œc.x/ h.x/xi (4.15)
While the learning rate, , and the difference between the real and the hypothe-
sized class label, Œc.x/ h.x/, have the same meaning and impact as before, what
has changed is the role of xi . In the case of boolean attributes, the value of xi decided
whether or not the weight should change. Here, however, it rather says how much
the weight should be affected: more in the case of higher attribute values.
The Multiplicative Rule In the case of WINNOW, too, essentially the same
learning formula can be used as in the binary-attributes case:
wi D wi ˛ c.x/h.x/ (4.16)
This said, one has to be careful about when to apply the formula. Previously, one
modified only the weights of attributes with xi D 1. Now that the attribute values
come from a continuous domain, some modification is needed. One possibility is
the following rule:
“Update weight wi only if the value of the i-th attribute is xi 0:5.”
Let us remark that both of these algorithms (perceptron learning and WINNOW)
usually find a relatively low-error solution even if the two classes are not linearly
separable—for instance, in the presence of noise.
Which Linear Classifier Is Better? At this point, however, another important
question needs to be discussed. Figure 4.8 shows three linear classifiers, each
perfectly separating the positive training examples from the negative. Knowing that
“good behavior” on the training set does not guarantee high performance in the
future, we have to ask: which of the three is likely to score best on future examples?
The Support Vector Machine Mathematicians who studied this problem found an
answer. When we take a look at Fig. 4.8, we can see that the dotted-line classifier
all but touches the nearest examples on either side; we say that its margin is small.
4.7 Numerical Domains and Support Vector Machines 85
Conversely, the margin is greater in the case of the solid-line classifier: the nearest
positive example on one side of the line, and the nearest example on the other of
the line are much farther than in the case of the other classifier. As it turns out, the
greater the margin, the higher the chances that the classifier will do well on future
data.
The technique of the support vector machines is illustrated in Fig. 4.9. The
solid line is the best classifier. The graph shows also two thinner lines, parallel
to the classifier, each at the same distance. The reader can see they pass through
the examples nearest to the classifier. These examples are called support vectors
(because, after all, each example is a vector of attributes).
The task for machine learning is to identify the support vectors that maximize the
margin. The concrete mechanisms for finding the optimum set of support vectors
exceed the ambition of an introductory text. The simplest technique would simply
try all possible n-tuples of examples, and measure the margin implied by each such
choice. This, however, is unrealistic in domains with many examples. Most of the
time, therefore, engineers rely on some of the many software packages available for
free on the internet.
• Can perceptron learning and WINNOW be used in numeric domains? How?
• Given that there are infinitely many linear classifiers capable of separating the
positive examples from the negative (assuming such separation exists), which of
them can be expected to give the best results on future data?
• What is a support vector? What is meant by the margin to be maximized?
Fig. 4.8 Linearly separable

classes can be separated in x2
infinitely many different
ways. Question is, which of
the classifiers that are perfect
on the training set will do best
on future data
x1
Fig. 4.9 The technique of the mar

support vector machine looks gin
for a separating hyperplane
that has the maximum margin
x2
x1
• Linear and polynomial classifiers define a decision surface that separates the
positive examples from the negative. Specifically, linear classifiers label the
examples according to the sign of the following expression:
w0 C w1 x1 C : : : wn xn
The concrete behavior is determined by the weights, wi . The task for machine
learning is to find appropriate values for these weights.
• The learning techniques from this chapter rely on the principle of “learning from
mistakes.” The training examples are presented, one by one, to the learner. Each
time the learner misclassifies an example, the weights are modified. When the
entire training set has been presented, one epoch has been completed. Usually,
several epochs are needed.
• Two weight-modification techniques were considered here: the additive rule of
perceptron learning, and the multiplicative rule of WINNOW.
• In domains with more than two classes, one can consider the use of a specialized
classifier for each class. A “master classifier” then chooses the class whose
classifier had the highest value of †niD0 wi xi .
• In domains with non-linear class boundaries, polynomial classifiers can some-
times be used. A second-order polynomial in a two-dimensional domain is
defined by the following expression:
w0 C w1 x1 C w2 x2 C w3 x12 C w4 x1 x2 C w5 x22
• The weights of the polynomial can be found by the same learning algorithms
as in the case of linear classifiers, provided that the non-linear terms (e.g., x1 x2 )
have been replaced (with the help of multipliers) by newly created attributes such
as z3 D x12 or z4 D x1 x2 .
• Polynomial classifiers are prone to overfit noisy training data. This is explained
by the excessive flexibility caused by the very high number of trainable weights.
• The potentially best class-separating hyperplane (among the infinitely many
candidates) is identified by the technique of the support vector machines (SVM)
that seek to maximize the distance of the nearest positive and the nearest negative
example from the hyperplane.
Historical Remarks The principle of perceptron learning was developed by
Rosenblatt [81], whereas WINNOW was proposed and analyzed by Littlestone [54].
The question of the capacity of linear and polynomial classifiers was analyzed by
Cover [18]. The principle of Support Vector Machines was invented by Vapnik [93]
as one of the consequences of the Computational Learning Theory which will be
the subject of Chap. 7.
Exercises
1. Write down equations for linear classifiers to implement the following func-
tions:
• At least two out of the boolean attributes x1 ; : : : ; x5 are true
• At least three out of the boolean attributes x1 ; : : : ; x6 are true, and at least one
of them is false.
2. Return to the examples from Table 4.2. Hand-simulate the perceptron learning
algorithm’s procedure, starting from a different initial set of weights than the one
used in the table. Try also a different learning rate.
3. Repeat the same exercise, this time using WINNOW. Do not forget to introduce
the additional “attributes” for what in perceptrons were the negative weights.
4. Write down the equation that defines a third-order polynomial in two dimensions.
How many multipliers (each with up to three inputs) would be needed if we
wanted to train the weights using the perceptron learning algorithm?
1. How can induction of linear classifiers be used to identify irrelevant attributes?

Hint: try to run the learning algorithm on different subsets of the attributes, and
then observe the error rate achieved after a fixed number of epochs.
2. Explain in what way it is true that the 1-NN classifier applied to a pair of
examples (one positive, the other negative) in a plane defines a linear classifier.
Invent a machine learning algorithm that uses this observation for the induction
of linear classifiers. Generalize the procedure to n-dimensional domains.
3. When is a linear classifier likely to lead to better classification performance on
independent testing examples than a polynomial classifier?
4. Sometimes, a linearly separable domain becomes linearly non-separable on
account of the class-label noise. Think of a technique capable of removing such
noisy examples. Hint: you may rely on an idea we have already encountered in
the field of k-NN classifiers.
1. Implement the perceptron learning algorithm and run it on the following

training set where six examples (three positive and three negative) are described
by four attributes:
x1 x2 x3 x4 Class
1 1 1 0 pos
0 0 0 0 pos
1 1 0 1 pos
1 1 0 0 neg
0 1 0 1 neg
0 0 0 1 neg
Observe that the linear classifier fails to reach zero error rate because the two
classes are not linearly separable.
2. Create a training set consisting of 20 examples described by five binary
attributes, x1 ; : : : ; x5 . Examples in which at least three attributes have values
xi D 1 are labeled as positive, all other examples are labeled as negative. Using
this training set as input, induce a linear classifier using perceptron learning.
Experiment with different values of the learning rate, . Plot a function where
the horizontal axis represents , and the vertical axis represents the number of
example-presentations needed for the classifier to correctly classify all training
examples. Discuss the results.
3. Use the same domain as in the previous assignment (five boolean attributes,
and the same definition of the positive class). Add to each example N additional
boolean attributes whose values are determined by a random-number generator.
Vary N from 1 to 20. Observe how the number of example-presentations needed
to achieve the zero error rate depends on N.
4. Again, use the same domain, but add attribute noise by changing the values of
randomly selected examples (while leaving class labels unchanged). Observe
what minimum error rate can then be achieved.
5. Repeat the last three assignments for different sizes of the training set,
evaluating the results on (always the same) testing set of examples that have
not been seen during learning.
6. Repeat the last four assignments, using WINNOW instead of the perceptron
learning. Compare the results in terms of the incurred computational costs.
These costs can be measured by the number of epochs needed to converge to
the zero error rate on the training set.
7. Define a domain with three numeric attributes with values from the unit interval,
Œ0; 1. Generate 100 training examples, labeling as positive those for which the
expression 1 x1 C x2 C x3 is positive. Use the “perceptron learning algorithm”
modified so that the following versions of the weight-updating rule are used:
(a) wi D wi C Œc.x/ h.x/xi
(b) wi D wi C Œc.x/ h.x/
P
(c) wi D wi C Œc.x/ P wi xi xi
(d) wi D wi C Œc.x/ wi xi
Will all of them converge to zero error rate on the training set? Compare the
speed of conversion.
8. Create a training set where each example is described by six boolean attributes,
xi ; : : : ; x6 . Label each example with one of the four classes defined as follows:
(a) C1 : at least five attributes have xi D 1.
(b) C2 : three or four attributes have xi D 1.
(c) C3 : two attributes have xi D 1.
(d) C4 : one attribute has xi D 1.
Use perceptron learning, applied in parallel to each of the four classes.
As a variation, use different numbers of irrelevant attributes, varying their
number from 0 to 20. See if the zero error rate on the training set can be
achieved.
Record the number of false positives and the number of false negatives observed
on an independent testing set.
Design an experiment showing that the performance of K binary classifiers,
connected in parallel as in Fig. 4.4, will decrease if we increase the number of
classes. How much is this observation pronounced in the presence of noise?
9. Run induction of linear classifiers on selected boolean domains from the UCI
repository3 and compare the results.
10. Experimenting with selected domains from the UCI repository, observe the
impact of the learning rate, , on the convergence speed of the perceptron
learning algorithm.
11. Compare the behavior of linear and polynomial classifiers. Observe how the
former wins in simple domains, and the latter in highly non-linear domains.
3
www.ics.uci.edu/˜mlearn/MLRepository.html.
Chapter 5
Artificial Neural Networks
Polynomial classifiers can model decision surfaces of any shape; and yet their
practical utility is limited because of the easiness with which they overfit noisy
training data, and because of the sometimes impractically high number of trainable
parameters. Much more popular are artificial neural networks where many simple
units, called neurons, are interconnected by weighted links into larger structures of
remarkably high performance.
The field of neural networks is too rich to be covered in the space we have
at our disposal. We will therefore provide only the basic information about two
popular types: multilayer perceptrons and radial-basis function networks. The
chapter describes how each of them classifies examples, and then describes some
elementary mechanisms to induce them from training data.
5.1 Multilayer Perceptrons as Classifiers
Throughout this chapter, we will suppose that all attributes are continuous. More-
over, it is practical (though not strictly necessary) to assume that they have been
normalized so that their values always fall in the interval Œ1; 1.
Neurons The function of a neuron, the basic unit of a multilayer perceptron, is quite
simple. A weighted sum of signals arriving at the input is subjected to a transfer
function. Several different transfer functions can be used; the one that is preferred
in this chapter is the so-called sigmoid, defined by the following formula where †
is the weighted sum of inputs:
1
f .†/ D (5.1)
1 C e†

DOI 10.1007/978-3-319-63913-0_5
92 5 Artificial Neural Networks
Fig. 5.1 A popular transfer

function: the sigmoid
f( Σwix i )
0.5
Σ wi x i
y1 yn output signals
... output neurons
... hidden neurons
input signals
x1 x2 x3 x4
Fig. 5.2 An example neural network consisting of two interconnected layers
Figure 5.1 shows the curve representing the transfer function. The reader can see
that f .†/ grows monotonically with the increasing value of † but is destined never
to leave the open interval .0; 1/ because f .1/ D 0, and f .1/ D 1. The vertical
axis is intersected at f(0)=0.5. We will assume that each neuron has the same transfer
function.
Multilayer Perceptron The neural network in Fig. 5.2 is known as the multilayer
perceptron. The neurons, represented by ovals, are arranged in the output layer
and the hidden layer.1 For simplicity, we will consider only networks with a single
hidden layer while remembering that it is quite common to employ two such layers,
even three, though rarely more than that. Another simplification is that we have
omitted the zero-th weights (providing for each neuron its trainable bias) so as not
to clutter the picture.
1
When we view the network from above, the hidden layer is obscured by the output layer.
5.1 Multilayer Perceptrons as Classifiers 93
While there is no communication between neurons of the same layer, adjacent

layers are fully interconnected. Importantly, each neuron-to-neuron link is associ-
ated with a weight. The weight of the link from the j-th hidden neuron to the i-th
.1/
output neuron is denoted as wji , and the weight of the link from the k-th attribute
.2/
to the j-th hidden neuron as wkj . Note that the first index always refers to the link’s
“beginning”; the second, to its “end.”
Forward Propagation When we present the network with an example, x D
.x1 ; : : : ; xn /, its attribute values are passed along the links to the neurons. The values
xk being multiplied by the weights associated with the links, the j-th hidden neuron
P .2/
receives as input the weighted sum, k wkj xk , and subjects this sum to the sigmoid,
P .2/
f . k wkj xk /. The i-th output neuron then receives the weighted sum of the values
coming from the hidden neurons and, again, subjects it to the transfer function. This
is how the i-th output is obtained. The process of propagating in this manner the
attribute values from the network’s input to its output is called forward propagation.
Each class is assigned its own output neuron, and the value returned by the i-th
output neuron is interpreted as the amount of evidence in support of the i-th class.
For instance, if the values obtained at three output neurons are y D .0:2; 0:6; 0:1/,
the classifier will label the given example with the second class because 0.6 is
greater than both 0.2 and 0.1.
In essence, this kind of two-layer perceptron calculates the following formula
where f is the sigmoid transfer function (see Eq. (5.1)) employed by the neurons;
.2/ .1/
wkj and wji are the links leading to the hidden and output layers, respectively, and
xk are the attribute values of the presented example:
X .1/
X .2/
yi D f . wji f . wkj xk // (5.2)
j k
A Numeric Example The principle of forward propagation is illustrated by the

numeric example in Table 5.1. At the beginning, the attribute vector x is presented.
Before reaching the neurons in the hidden layer, the attribute values are multiplied
by the corresponding weights, and the weighted sums are passed on to the sigmoid
functions. The results (h1 D 0:32 and h2 D 0:54) are then multiplied by the next
layer of weights, and forwarded to the output neurons where they are again subjected
to the sigmoid function. This is how the two output values, y1 D 0:66 and y2 D 0:45,
have been obtained. The evidence supporting the class of the “left” output neuron
is higher than the evidence supporting the class of the “right” output neuron. The
classifier therefore chooses the left neuron’s class.
Universal Classifier Mathematicians have been able to prove that, with the right
choice of weights, and with the right number of the hidden neurons, Eq. (5.2) can
approximate with arbitrary accuracy any realistic function. The consequence of this
so-called universality theorem is that the multilayer perceptron can in principle be
used to address just about any classification problem. What the theorem does not
Table 5.1 Example of forward propagation in a multilayer perceptron

Task. Forward-propagate x D .x1 ; x2 / D .0:8; 0:1/ through the network below.
y1=0.66 y2=0.45
0.9 −0.1
−0.3 0.5
h1=0.32 h2=0.54
−1.0 0.7
0.1 0.5
x1=0.8 x2=0.1
Solution.
inputs of hidden-layer neurons:
.2/
z1 D 0:8 .1:0/ C 0:1 0:5 D 0:75
.2/
z2 D 0:8 0:1 C 0:1 0:7 D 0:15
outputs of hidden-layer neurons:
.2/ 1
h1 D f .z1 / D 1Ce.0:75/
D 0:32
.2/ 1
h2 D f .z2 / D 1Ce0:15
D 0:54
inputs of output-layer neurons:
.1/
z1 D 0:32 0:9 C 0:54 0:5 D 0:56
.1/
z2 D 0:32 .0:3/ C 0:54 .0:1/ D 0:15
outputs of output-layer neurons:
.1/ 1
y1 D f .z1 / D 1Ce0:56
D 0:66
.1/ 1
y2 D f .z2 / D 1Ce.0:15/
D 0:45
5.2 Neural Network’s Error 95
tell us, though, is how many hidden neurons are needed, and what the individual
weight values should be. In other words, we know that the solution exists, yet there
is no guarantee we will ever find it.
• Explain how the example described by a vector of continuous attributes is
forward-propagated through the multilayer perceptron. How is the network’s
output interpreted?
• What is the transfer function? Write down the formula defining the sigmoid
transfer function, and describe its shape.
• What is the universality theorem? What does it tell us, and what does it not tell
us?
5.2 Neural Network’s Error
Let us defer to a later chapter the explanation of a technique to train the multilayer
perceptron (to find its weights). Before we can address this issue, it is necessary to
prepare the soil by taking a closer look at the method of classification, and at the
way the accuracy of this classification is evaluated.
Error Rate Suppose a training example, x, with known class, c.x/, has been pre-
sented to an existing version of the multilayer perceptron. The forward-propagation
step establishes the label, h.x/. If h.x/ ¤ c.x/, an error has been made. This may
happen to some other examples, too, and we want to know how often this happens.
We want to know the error rate—which is for a given set of examples obtained
by dividing the number of errors by the number of examples. For instance, if the
classifier misclassifies 30 out of 200 examples, the error rate is 30=200 D 0:15.
This, however, fails to give the full picture of the network’s classification
performance. What the error rate neglects to reflect is the sigmoid function’s ability
to measure the size of each error.
An example will clarify the point. Suppose we have two different networks to
choose from, each with three output neurons corresponding to classes denoted by
C1 ; C2 ; and C3 . Let us assume that, for some example x, the first network outputs
y1.x/ D .0:5; 0:2; 0:9/ and the second, y2.x/ D .0:6; 0:6; 0:7/. This means that
both will label x with the third class, h1.x/ D h2.x/ D C3 . If the correct answer
is c.x/ D C2 , both have erred, but the error does not appear to be the same. The
reader will have noticed that the first network was “very sure” about the class being
C3 (because 0.9 is clearly greater than the other two outputs, 0.5 and 0.2), whereas
the second network was less certain, the differences of the output values (0.6, 0.6,
and 0.7) being so small as to give rise to the suspicion that C3 has won merely by
chance. Due to its weaker commitment to the incorrect class, the second network is
somehow less wrong than the first.
This is the circumstance that can be captured by a more appropriate error
function, the mean square error (MSE).
Target Vector Before proceeding to the definition of the mean square error,
however, we must introduce yet another important concept, the target vector which,
too, depends on the concrete example, x. Let us denote it by t.x/. In a domain with
m classes, the target vector, t.x/ D .t1 .x/; : : : ; tm .x//, consists of m binary numbers.
If the example belongs to the i-th class, then ti .x/ D 1 and all other elements in
this vector are tj .x/ D 0 (where j ¤ i). For instance, suppose the existence of three
different classes, C1 ; C2 ; and C3 , and let x be known to belong to C2 . In the ideal
case, the second neuron should output 1, and the two other neurons should output
0.2 The target is therefore t.x/ D .t1 ; t2 ; t3 / D .0; 1; 0/.
Mean Square Error The mean square error is defined using the differences
between the elements of the output vector and the target vector:
1X
m
MSE D .ti yi /2 (5.3)
m iD1
When calculating the network’s MSE, we have to establish for each output neuron
the difference between its output and the corresponding element of the target vector.
Note that the terms in the parentheses, .ti yi /, are squared to make sure that negative
differences are not subtracted from positive ones.
Returning to the example of the two networks mentioned above, if the target
vector is t.x/ D .0; 1; 0/, then these are the mean square errors:
MSE1 D 13 Œ.0 0:5/2 C .1 0:2/2 C .0 0:9/2 / D 0:57
MSE2 D 13 Œ.0 0:6/2 C .1 0:6/2 C .0 0:7/2 D 0:34
As expected, MSE2 <MSE1 , which is in line with our intuition that the second
network is “less wrong” on x than the first network.
2
More precisely, the outputs will only approach 1 and 0 because the sigmoid function is bounded
by the open interval, .0; 1/.
5.3 Backpropagation of Error 97
• In what sense do we say that the traditional error rate does not provide enough
information about a neural network’s classification accuracy?
• Explain the difference between a neural network’s output, the example’s class,
and the target vector.
• Write down the formulas defining the error rate and the mean square error.
5.3 Backpropagation of Error
In the multilayer perceptron, the parameters to affect the network’s behavior are the
.1/ .2/
sets of weights, wji and wkj . The task for machine learning is to find for these
weights such values that will optimize the network’s classification performance.
Just like in the case of linear classifiers, this is achieved by training. This section is
devoted to one popular technique capable of accomplishing this task.
The General Scenario In principle, the procedure is the same as in the previous
chapter. At the beginning, the weights are initialized to small random numbers, typ-
ically from the interval .0:1; 0:1/. After this, the training examples are presented,
one by one, and each of them is forward-propagated to the network’s output. The
discrepancy between this output and the example’s target vector then tells us how
to modify the weights (see below). After the weight modification, the next example
is presented. When the last training example has been reached, one epoch has been
completed. In multilayer perceptrons, the number of epochs needed for successful
training is much greater than in the case of linear classifiers: it can be thousands,
tens of thousands, even more.
The Gradient Descent Before we proceed to the concrete formulas for weight
adjustment, we need to develop a better understanding of the problem’s nature.
Figure 5.3 will help us. The vertical axis represents the mean square error, expressed
as a function of the network’s weights (plotted along the horizontal axes). For
graphical convenience, we assume that there are only two weights. This, of course,
is unrealistic to say the least. But if we want an instructive example, we simply
cannot afford more dimensions—we can hardly visualize ten dimensions, can we?
The message we want to drive home at this point is that the error function can be
imagined as a kind of a “landscape” whose “valleys” represent the function’s local
minima. The deepest of them is the global minimum, and this is what the training
procedure is, in the ideal case, expected to reach; more specifically, it should find
the set of weights corresponding to the global minimum.
A quick glance at Fig. 5.3 tells us that any pair of weights defines for the given
training example a concrete location on the landscape, typically somewhere on one
of the slopes. Any weight change will then result in different coordinates along the
horizontal axes, and thus a different location on the error function. Where exactly
this new location is, whether “up” or “down” the slope, will depend on how much,
and in what direction, each of the weights has changed. For instance, it may be that
increasing both w1 and w2 by 0.1 will lead only to a minor reduction of the mean
square error; whereas increasing w1 by 0.3 and w2 by 0.1 will reduce it considerably.
Fig. 5.3 For a given example, each set of weights implies a certain mean square error. Training
should reduce this error as quickly as possible
In the technique discussed here, we want weight changes that will bring about the
steepest descent along the error function. Recalling the terminology from Chap. 1,
this is a job for hill-climbing search. The best-known technique used to this end in
multilayer perceptrons is backpropagation of error.
Backpropagation of Error The specific weight-adjusting formulas can be derived
from Eq. (5.2) by finding the function’s gradient. However, as this book is meant
for practitioners, and not for mathematicians, we will skip the derivation, and focus
instead on explaining the learning procedure’s behavior.
To begin with, it is reasonable to assume that the individual neurons differ in
their contributions to the overall error. Some of them “spoil the game” more than the
others. If this is the case, the reader will agree that the links leading to these neurons
should undergo greater weight changes than the links leading to less offending
neurons.
Fortunately, each neuron’s amount of “responsibility” for the overall error can
be easily obtained. Generally speaking, the concrete choice of formulas depends on
what transfer function has been used. In the case of the sigmoid (see Eq. (5.1)),
the responsibility is calculated as follows:
5.3 Backpropagation of Error 99
.1/
Output-layer neurons: ıi D yi .1 yi /.ti yi /
Here, .ti yi / is the difference between the i-th output and the corresponding
target value. This difference is multiplied by yi .1 yi /, a term whose minimum
value is reached when y1 D 0 or yi D 1 (a “strong opinion” as to whether x should
or should not be labeled with the i-th class); the term is maximized when yi D 0:5, in
.1/
which case the “opinion” can be deemed neutral. Note that the sign of ıi depends
only on .ti yi / because yi .1 yi / is always positive.
.2/ P .1/
Hidden-layer neurons: ıj D hj .1 hj / i ıi wji
The responsibilities of the hidden neurons are calculated by “backpropagating”
the output-neurons’ responsibilities obtained in the previous step. This is the role of
P .1/ .1/
the term i ıi wji . Note that each ıi (the responsibility of the i-th output neuron)
is multiplied by the weight of the link connecting the i-th output neurons to the j-th
hidden neuron. The weighted sum is multiplied by, hj .1 hj /, essentially the same
term as the one used in the previous step, except that hj has taken the place of yi .
Weight Updates Now that we know the responsibilities of the individual neurons,
we are ready to update the weights of the links leading to them. Similarly to
perceptron learning, an additive rule is used:
.1/ .1/ .1/
output-layer neurons: wji WD wji C ıi hj
.2/ .2/ .2/
hidden-layer neurons: wkj WD wkj C ıj xk
.1/ .2/
The extent of weight correction is therefore determined by ıi hj or ıj xk .
.1/ .2/
Two observations can be made. First, the neurons’ responsibilities, ıi or ıj , are
multiplied by , the learning rate which, theoretically speaking, should be selected
from the unit interval, 2 .0; 1/; however, practical implementations usually rely
on smaller values, typically less than 0:1. Second, the terms are also multiplied by
hj 2 .0; 1/ and xk 2 Œ0; 1, respectively. The correction is therefore quite small, but
its effect is relative. If the added term’s value is, say, 0.02, then smaller weights,
.1/
such as wij D 0:01, will be affected more significantly than greater weights such
.1/
as wij D 1:8.
The whole training procedure is summarized by the pseudocode in Table 5.2. The
reader will benefit also from taking a closer look at the example given in Table 5.3
that provides all the necessary details of how the weights are updated in response to
one training example.
• Why is the training technique called “backpropagation of error”? What is the
rationale behind the need to establish the “neurons’ responsibilities”?
Table 5.2 Backpropagation of error in a neural network with one hidden layer
1. Present example x to the input layer and propagate it through the network.
2. Let y D .y1 ; : : : ym / be the output vector, and let t.x/ D .t1 ; : : : tm / be the target vector.
.1/
3. For each output neuron, calculate its responsibility, ıi , for the network’s error:
.1/
ıi D yi .1 yi /.ti yi /
.2/
4. For each hidden neuron, calculate its responsibility, ıj , for the network’s error. While doing
.1/
so, use the responsibilities, ıi , of the output neurons as obtained in the previous step.
.2/ P .1/
ıj D hj .1 hj / i ıi wji
5. Update the weights using the following formulas, where is the learning rate:
.1/ .1/ .1/
output layer: wji WD wji C ıi hj ; hj : the output of the j-th hidden neuron
.2/ .2/ .2/
hidden layer: wkj WD wkj C ıj xk ; xk : the value of the k-th attribute
6. Unless a termination criterion has been satisfied, return to step 1.
• Discuss the behaviors of the formulas for the calculation of the responsibilities
of the neurons in the individual layers.
• Explain the behaviors of the formulas used to update the weights. Mention some
critical aspects of these formulas.
5.4 Special Aspects of Multilayer Perceptrons
Limited space prevents us from the detailed investigation of the many features that
make the training of multilayer perceptrons more art than science. To do justice to
all of them, another chapter at least the size of this one would be needed. Still, the
knowledge of certain critical aspects is vital if the training is ever to succeed. Let us
therefore briefly survey some of the more important ones.
Computational Costs Backpropagation of error is computationally expensive.
Upon the presentation of an example, the responsibility of each individual neuron
has to be calculated, and the weights then modified accordingly. This has to be
repeated for all training examples, usually for many epochs. To get an idea of the
real costs of all this, consider a network that is to classify examples described
by 100 attributes, a fairly realistic case. If there are 100 hidden neurons, then the
number of weights in this lower layer is 100 100 D 104 . This then is the number
of weight changes after each training example. Note that the upper-layer weights
can be neglected, in these calculations, as long as the number of classes is small.
For instance, in a domain with three classes, the number of upper-layer weights is
100 3 D 300 which is much less than 104 .
5.4 Special Aspects of Multilayer Perceptrons 101
Suppose the training set consists of 105 examples, and suppose that the training
will continue for 104 epochs. In this event, the number of weight-updates will be
104 105 104 D 1013 . This looks like a whole lot, but many applications are even
more demanding. Some fairly ingenious methods to make the training more efficient
have therefore been developed. These, however, are outside the scope of our interest
here.
Target Values Revisited For simplicity, we have so far assumed that each target
value is either 1 or 0. This may not be the best choice. For one thing, these values
Table 5.3 Illustration of the backpropagation of error

Task. In the neural network below, let the transfer function be f .†/ D 1Ce1† . Using
backpropagation of error (with D 0:1), show how the weights are modified after the
presentation of the following example: Œx; t.x/ D Œ.1; 1/; .1; 0/
y1 = 0.65 y2 = 0.59
(1) (1)
1 δ1 −1 2 δ2
1 1 1
(2) h1 = 0.12 h2 = 0.5

δ1 1 1 2 δ (2)
1 2
−1
1
x1 x2
Forward propagation.
The picture shows the state after forward propagation when the signals leaving the hidden and
the output neurons have been calculated as follows:
1
• h1 D 1Ce.2/
D 0:12
1
h2 D 1Ce0
D 0:5
1
y1 D 1Ce.0:12C0:5/
D 0:65
1
y2 D 1Ce.0:12C0:5/
D 0:59
(the solution continues on the next page)

(continued)
Table 5.3 (continued)

Backpropagation of error (cont. from the previous page)
The target vector being t.x/ D .1; 0/, and the output vector y D .0:65; 0:59/, the task is to
establish each neuron’s responsibility for the output error. Here are the calculations for the output
neurons:
.1/
• ı1 D y1 .1 y1 /.t1 y1 / D 0:65.1 0:65/.1 0:65/ D 0:0796
.1/
ı2 D y2 .1 y2 /.t2 y2 / D 0:59.1 0:59/.0 0:59/ D 0:1427
Using these values, we calculate the responsibilities of the hidden neurons. Note that we will
P .1/ .1/
first calculate (and denote by 1 and 2 ) the weighted sums, i ıi wij , for each of the two
hidden neurons.
.1/ .1/ .1/ .1/
• 1 D ı1 w11 C ı2 w12 D 0:0796 1 C .0:1427/ .1/ D 0:2223
.1/ .1/ .1/ .1/
2 D ı1 w21 C ı2 w22 D 0:0796 1 C .0:1427/ 1 D 0:0631
.2/
ı1 D h1 .1 h1 /1 D 0:12.1 0:12/ 0:2223 D 0:0235
.2/
ı2 D h2 .1 h2 /2 D 0:5.1 0:5/ .0:0631/ D 0:0158
Once the responsibilities are known, the weight modifications are straightforward:
.1/ .1/ .1/
• w11 D w11 C ı1 h1 D 1 C 0:1 0:0796 0:12 D 1:00096
.1/ .1/ .1/
w21 D w21 C ı1 h2 D 1 C 0:1 0:0796 0:5 D 1:00398
.1/ .1/ .1/
w12 D w12 C ı2 h1 D 1 C 0:1 .0:1427/ 0:12 D 1:0017
.1/ .1/ .1/
w22 D w22 C ı2 h2 D 1 C 0:1 .0:1427/ 0:5 D 0:9929
.2/ .2/ .2/
• w11 D w11 C ı1 x1 D 1 C 0:1 .0:0235/ 1 D 1:0024
.2/ .2/ .2/
w21 D w21 C ı1 x2 D 1 C 0:1 .0:0235/ .1/ D 1:0024
.2/ .2/ .2/
w12 D w12 C ı2 x1 D 1 C 0:1 0:0158 1 D 1:0016
.2/ .2/ .2/
w22 D w22 C ı2 x2 D 1 C 0:1 0:0158 .1/ D 0:9984
The weights having been updated, the network is ready for the next example.
can never be reached by a neuron’s output, yi . Moreover, the weight changes in the
vicinity of these two extremes are miniscule because the calculation of the output-
.1/
neuron’s responsibility, ıi D yi .1 yi /.ti yi /, returns a value very close to zero
whenever yi approaches 0 or 1. Finally, we know that the classifier chooses the class
whose output neuron has returned the highest value. The individual neuron’s output
precision therefore does not matter much; more important is the comparison with
the other outputs. If the forward propagation results in y D .0:9; 0:1; 0:2/, then the
example is bound to be labeled with the first class (the one supported by yi D 0:9),
and this decision will not be swayed by minor weight changes.
In view of these arguments, more appropriate values for the target are recom-
mended: for instance, ti .x/ D 0:8 if the example belongs to the i-th class, and ti .x/ D
0:2 if it does not. Suppose there are three classes, C1 ; C2 , and C3 , and suppose
c.x/ D C1 . In this case, the target vector will be defined as t.x/ D .0:8; 0:2; 0:2/.
Both 0.8 and 0.2 find themselves in regions of relatively high sensitivity of the
sigmoid function, and as such will eliminate most of the concerns raised in the
previous paragraph.
5.4 Special Aspects of Multilayer Perceptrons 103
Local Minima Figure 5.3 illustrated the main drawback of the gradient-descent
approach when adopted by multilayer perceptron training. The weights are changed
in a way that guarantees descent along the steepest slope. But once the bottom of a
local minimum has been reached, there is nowhere else to go—which is awkward:
after all, the ultimate goal is to reach the global minimum. Two things are needed
here: first, a mechanism to recognize the local minimum; second, a method to
recover from having fallen into one.
One possibility of timely identification of local minima during training is to keep
track of the mean square error, and to sum it up over the entire training set at the end
of each epoch. Under normal circumstances, this sum tends to go down from one
epoch to another. Once it seems to have reached a plateau where hardly any error
reduction can be observed, the learning process is suspected of being trapped in a
local minimum.
Techniques to overcome this difficulty usually rely on adaptive learning rates (see
below), and on adding new hidden neurons (see Sect. 5.5). Generally speaking, the
problem is less critical in networks with many hidden neurons. Also, local minima
tend to be shallower, and less frequent, if all weights are very small, say, from the
interval .0:01; 0:01/.
Adaptive Learning Rate While describing backpropagation of error, we assumed
that the user-set learning rate, , was a constant. This, however, is rarely the case
in realistic applications. Very often, the training starts with a large which then
gradually decreases in time. The motivation is easy to guess. At the beginning,
greater weight changes reduce the number of epochs, and they may even help the
learner to “jump over” some local minima. Later on, however, this large might
lead to “overshooting” the global minimum, and this is why its value should be
decreased. If we express the learning rate as a function of time, .t/, where t tells
us which epoch the training is now going through, then the following negative-
exponential formula will gradually reduce the learning rate (˛ is the slope of the
negative exponential, and .0/ is the learning rate’s initial value):
.t/ D .0/e˛t (5.4)
It should perhaps be noted that some advanced weight-changing formulas are

capable of reflecting “current tendencies.” For instance, it is quite popular to
implement “momentum”: if the last two weight changes were both positive (or
both negative), it makes sense to increase the weight-changing step; if, conversely,
a positive change was followed by a negative change (of vice versa), the step should
be reduced so as to prevent overshooting.
Overtraining Multilayer perceptrons share with polynomial classifiers one
unpleasant property. Theoretically speaking, they are capable of modeling any
decision surface, and this makes them prone to overfitting the training data. The
reader remembers that overfitting typically means perfect classification of noisy
training examples, which is inevitably followed by disappointing performance in
the future.
For small multilayer perceptrons, this problem is not as painful; they are not
flexible enough to overfit. But as the number of hidden neurons increases, the
network gains in flexibility, and overfitting can become a real concern. However,
as we will learn in the next section, this does not mean that we should always prefer
small networks!
There is a simple way to discover whether the training process has reached the
“overfitting” stage. If the training set is big enough, we can afford to leave aside
some 10–20% examples. These will never be used for backpropagation of error;
rather, after each epoch, the sum of mean square errors on these withheld examples
is calculated. At the beginning, the sum will tend to go down, but only up to a
certain moment; then it starts growing again, alerting the engineer that the training
has begun to overfit the data.
• What do you know about the computational costs of the technique of backprop-
agation of error?
• Explain why this section recommended the values of the target vector to be
chosen from {0.8, 0.2} instead of from {1,0}.
• Discuss the problem of local minima. Why do they represent such a problem for
training? How can we reduce the danger of getting trapped in one?
• What are the benefits of an adaptive learning rate? What formula has been
recommended for it?
• What do you know about the danger that the training of a neural network might
result in overfitting the training data?
5.5 Architectural Issues
So far, one important question has been neglected: how many hidden neurons to
use? If there are only one or two, the network will lack flexibility; not only will it
be unable to model a complicated decision surface; the training of this network will
be prone to get stuck in a local minimum. At the other extreme, using thousands of
neurons will not only increase computational costs because of the need to train so
many neurons. The network will be more flexible than needed. As a result, it will
easily overfit the data. As usual, in situations of this kind, some compromise has to
be found.
Performance Versus Size Suppose you decide to run the following experiment.
The available set of pre-classified examples is divided into two parts, one for
training, the other for testing. Training is applied to several neural networks, each
5.5 Architectural Issues 105
Fig. 5.4 The error rate error rate on testing data

measured on testing examples
depends on the number of
0.4
neurons in the hidden layer
0.3
0.2
0.1
number of hidden neurons
with a different number of hidden neurons. The networks are trained until no
reduction of the training-set error rate is observed. After this, the error rate on the
testing data is measured.
Optimum Number of Neurons When plotted in a graph, the results will typically
look something like the situation depicted in Fig. 5.4. Here, the horizontal axis
represents the number of hidden neurons; the vertical axis, the error rate measured
on the testing set. Typically, the error rate will be high in the case of very small
networks because these lack adequate flexibility, and also suffer from the danger
of getting stuck in local minima. These two weaknesses can be mitigated if we
increase the number of hidden neurons. As shown in the graph, the larger networks
then exhibit lower error rates. But then, networks that are too large are vulnerable
to overfitting. This is why, after a certain point, the testing-set error starts growing
again (the right tail of the graph).
The precise shape of the curve depends on the complexity of the training data.
Sometimes, the error rate is minimized when the network contains no more than
3–5 hidden neurons. In other domains, however, the minimum is reached only when
hundreds of hidden neurons are used. Yet another case worth mentioning is the
situation where the training examples are completely noise-free. In a domain of this
kind, overfitting may never become an issue, and the curve’s right tail may not grow
at all.
Search for Appropriate Size The scenario described above is too expensive to be
employed in practical applications. After all, we have no idea whether we will need
just a few neurons, or dozens of them, or hundreds, and we may have to re-run the
computationally intensive training algorithm a great many times before being able
to establish the ideal size. Instead, we would like to have at our disposal a technique
capable of finding the appropriate size more efficiently.
One such technique is summarized by the pseudocode in Table 5.4. The idea is to
start with a very small network that only has a few hidden neurons. After each epoch,
the learning algorithm checks the sum of the mean square errors observed on the
training set. This sum of errors is likely to keep decreasing with the growing number
of epochs—but only up to a certain point. When this is reached, the network’s
performance no longer improves, either because of its insufficient flexibility that
makes correct classification impossible, or because it “fell” into a local minimum.
Fig. 5.5 When the MSE

training-set mean square error
(MSE) does not seem to go
down, further improvement
can be achieved by adding
new hidden neurons. In this
particular case, this has been
done twice
number of epochs
When this is observed, a few more neurons with randomly initialized weights are
added, and the training is resumed.
Usually, the added flexibility makes further error reduction possible. In the
illustration from Fig. 5.5, the error rate “stalled” on two occasions; adding new
hidden neurons then provided the necessary new flexibility.
• Discuss the shape of the curve from Fig. 5.4. What are the shortcomings of very
small and very large networks?
• Explain the algorithm that searches for a reasonable size of the hidden layer.
What main difficulties does it face?
5.6 Radial-Basis Function Networks
The behavior of an output neuron in multilayer perceptrons is similar to that of

a linear classifier. As such, it may be expected to fare poorly in domains with
Table 5.4 Gradual search for a good size of the hidden layer
1. At the beginning, use only a few hidden neurons, say, five.
2. Train the network until the mean square error no longer seems to improve.
3. At this moment, add a few (say, three) neurons to the hidden layer, each with randomly
initialized weights, and resume training.
4. Repeat the previous two steps until a termination criterion has been satisfied. For
instance, this can be the case when the new addition does not result in a significant
error reduction, or when the hidden layer exceeds a user-set maximum size.
5.6 Radial-Basis Function Networks 107
y1 yK
... output layer
ϕ ϕm
1 1 ϕ2
...
gaussian functions
x1 x2
Fig. 5.6 Radial-basis function network
classes that are not linearly separable. In the context of the neural networks,
however, this limitation is not necessarily hurtful. Thing is, the original examples
have been transformed by the sigmoid functions in the hidden layer. Consequently,
the neurons in the output layer deal with new “attributes,” those obtained by this
transformation. In the process of training, these transformed examples (the outputs
of the hidden neurons) may become linearly separable so that the output-layer
neurons can separate the two classes without difficulties.
The Alternative There is another way of transforming the attribute values; by
using the so-called radial-basis f unction, RBR, as the transfer function employed
by the hidden-layer neurons. This is the case of the network depicted in Fig. 5.6. An
example presented to the input is passed through a set of neurons that each return a
value denoted as 'j .
Radial-Basis Function, RBF In essence, this is based on the normal distribution
that we already know from Chap. 2. Suppose that the attributes describing the
examples all fall into some reasonably sized interval, say Œ1; 1. For a given
variance, 2 , the following equation defines the n-dimensional gaussian surface
centered at j D Œj1 ; : : : jn :
†niD1 .xi ji /2

'j .x/ D expf g (5.5)
2 2
In a sense, 'j .x/ measures similarity between the example vector, x, and the
gaussian center, j : the larger the distance between the two, the smaller the value
of 'j .x/. If x is to be classified, the network first redescribes it as '.x/ D
Œ'1 .x/; : : P
: ; 'm .x/. The output signal of the i-th output neuron is then calculated
m
as yi D jD0 wji 'j .x/, where wji is the weight of the link from the j-th hidden
neuron to the i-th output neuron (the weights w0i are connected to a fixed '0 D 1).
This output signal being interpreted as the amount of evidence supporting the i-th
class, the example is labeled with the i-th class if yi D maxk .yk /.
Output-Layer Weights It is relatively simple to establish the output-layer weights,
wij . Since there is only one layer of weights to be trained, we can just as well rely
on the perceptron learning algorithm described in Chap. 4, applying it to examples
whose descriptions have been transformed by the RBF functions in the hidden-layer
neurons.
Gaussian Centers It is a common practice to identify the gaussian centers, j ,
with the individual training examples. If the training set is small, we can simply
use one hidden neuron per training example. In many realistic applications, though,
the training sets are much bigger, which can mean thousands of hidden neurons,
or even more. Realizing that such large networks are unwieldy, many engineers
prefer to select for the centers only a small subset of the training examples. Very
often, a random choice is enough. Another possibility to identify groups of “similar”
vectors, and then use for each RBF neuron the center of one group. To identify such
groups of similar vectors is a task for so-called cluster analysis that will discussed
in great detail in Chap. 14.
RBF-Based Support Vector Machines The RBF neurons transform the original
example description into a new vector that consists of the values, 1 ; : : : m . Most
of the time, this transformation increases the chances that the examples thus
transformed will be linearly separable. This makes it possible to subject them to
a linear classifier whose weights are trained by perceptron learning.
This is perhaps an appropriate place to mention that it is also quite popular to
apply to the transformed examples the idea of the support vector machine introduced
in Sect. 4.7. In this event, the resulting machine-learning tool is usually referred to
as RBF-based SVM. Especially in domains where the boundary between the classes
is highly non-linear, this classifier is more powerful than the plain (linear) SVM.
Computational Costs RBF-network training consists of two steps. First, the
centers of the radial-basis functions (the Gaussians) are selected. Second, the output-
neurons’ weights are obtained by training. Since there is only one layer to train,
the process is computationally much less intensive than the training in multilayer
perceptrons.
• Explain the principle of the radial-basis function network. In what aspects does
it differ from the multilayer perceptron?
• How many weights need to be trained in a radial-basis function network? How

can the training be accomplished?
• What are the possibilities for the construction of the gaussian hidden layer?
• The basic unit of a multilayer perceptron is a neuron. The neuron accepts a

weighted sum of inputs, and subjects this sum to a transfer function. Several
different transfer functions can be used. The one chosen in this chapter is the
sigmoid defined by the following equation where † is the weighted sum of
inputs:
1
f .†/ D
1 C e†
Usually, all neurons use the same transfer function.

• The simple version of the multilayer perceptron described in this chapter consists
of one output layer and one hidden layer of neurons. Neurons in adjacent layers
are fully interconnected; but there are no connections between neurons in the
same layer. An example presented to the network’s input is forward-propagated
to its output, implementing, in principle, the following function:
X .1/
X .2/
yi D f . wji f . wkj xk //
j k
.1/ .2/
Here, wji and wkj are the weights of the output and the hidden neurons,
respectively, and f is the sigmoid function.
• The training of multilayer perceptron is accomplished by a technique known
as backpropagation of error. For each training example, the technique first
establishes each neuron’s individual responsibility for the network’s overall error,
and then updates the weights according to these responsibilities.
Here is how the responsibilities are calculated:
.1/
output neurons: ıi D yi .1 yi /.ti yi /
.2/ P .1/
hidden neurons: ıj D hj .1 hj / i ıi wij
Here is how the weights are updated:

.1/ .1/ .1/
output layer: wji WD wji C ıi hj
.2/ .2/ .2/
hidden layer: wkj WD wkj C ıj xk
• Certain aspects of backpropagation by error have been discussed here. Among

these, the most important are computational costs, the existence of local minima,
adaptive learning rate, the danger of overfitting, and the problems of how to
determine the size of the hidden layer.
• An alternative is the radial-basis function (RBF) network. For the transfer
function at the hidden-layer neurons, the gaussian function is used. The output-
layer neurons often use the step function (in principle, the linear classifier), or
simply a linear function of the inputs.
• In RBF networks, each gaussian center is identified with one training example. If
there are too many such examples, a random choice can be made. The gaussian’s
standard deviation is in this simple version set to 2 D 1,
• The output-layer neurons in RBF networks can be trained by perceptrons
learning. Only one layer of weights needs to be trained.
• Sometimes, the idea of support vector machine (SVM) is applied to the outputs
of hidden neurons. The resulting tool is then called the RBF-based SVM.
Historical Remarks Research of neural networks was famously delayed by the
skeptical views formulated by Minsky and Papert [66]. The pessimism expressed by
such famous authors was probably the main reason why an early version of neural-
network training by Bryson and Ho [13] was largely overlooked, a fate soon to
be shared by an independent successful attempt by Werbos [95]. It was only after
the publication of the groundbreaking volumes by Rumelhart and McClelland [83],
where the algorithm was independently re-invented, that the field of artificial neural
networks became established as a respectable scientific discipline. The gradual
growth of the multilayer perceptron was proposed by Ash [1]. The idea of radial-
basis functions was first cast in the neural-network setting by Broomhead and
Lowe [12].
Exercises
1. Return to the illustration of backpropagation of error in Table 5.3. Using only a

pen, paper, and calculator, repeat the calculations for a slightly different training
example: x D .1; 1/; t.x/ D .0; 1/.
2. Hand-simulating backpropagation of error as in the previous example, repeat the

calculation for the following two cases:
.1/ .1/ .1/ .1/
High output-layer weights: w11 D 3:0; w12 D 3:0; w21 D 3:0; w22 D 3:0
.1/ .1/ .1/ .1/
Small output-layer weights: w11 D 0:3; w12 D 0:3; w21 D 0:3; w22 D 0:3
Observe the relative changes in the weights in each case.
3. Consider a training set containing of 105 examples described by 1000 attributes.
What will be the computational costs of training a multilayer perceptron with
1000 hidden neurons and ten output neurons for 105 epochs?
1. How will you generalize the technique of backpropagation of error so that it can
be used in a multilayer perceptron with more that one hidden layer?
2. Section 5.1 suggested that all attributes should be normalized here to the interval
Œ1:0; 1:0. How will the network’s classification and training be affected if the
attributes are not so normalized? (hint: this has something to do with the sigmoid
function).
3. Discuss the similarities and differences of the classification procedures used in
radial-basis functions and in multilayer perceptrons.
4. Compare the advantages and disadvantages of radial-basis function networks in
comparison with multilayer perceptrons.
1. Write a program that implements backpropagation of error for a predefined

number of output and hidden neurons. Use a fixed learning rate, .
2. Apply the program implemented in the previous task to some benchmark
domains from the UCI repository.3 Experiment with different values of , and
see how they affect the speed of convergence.
3. For a given data set, experiment with different numbers of hidden neurons in the
multilayer perceptron, and observe how they affect the network’s ability to learn.
4. Again, experiment with different numbers of hidden neurons. This time, focus
on computational costs. How many epochs are needed before the network
converges? Look also at the evolution of the error rate.
5. Write a program that will, for a given training set, create a radial-basis function
network. For large training sets, select randomly the examples that will define
the gaussian centers.
6. Apply the program implemented in the previous task to some benchmark
domains from the UCI repository.
3
Chapter 6
Decision Trees
The classifiers discussed in the previous chapters expect all attribute values to be
presented at the same time. Such a scenario, however, has its flaws. Thus a physician
seeking to come to grips with the nature of her patient’s condition often has nothing
to begin with save a few subjective symptoms. And so, to narrow the field of
diagnoses, she prescribes lab tests, and, based on the results, perhaps other tests
still. At any given moment, then, the doctor considers only “attributes” that promise
to add meaningfully to her current information or understanding. It would be absurd
to ask for all possible lab tests (thousands and thousands of them) right from the
start.
The lesson is, exhaustive information often is not immediately available; it may
not even be needed. The classifier may do better choosing the attributes one at a
time, according to the demands of the situation. The most popular tool targeting this
scenario is a decision tree. The chapter explains its classification behavior, and then
presents a simple technique that induces a decision tree from data. The reader will
learn how to benefit from tree-pruning, and how to convert the induced tree to a set
of rules.
6.1 Decision Trees as Classifiers
The training set shown in Table 6.1 consists of eight examples described by three
attributes and labeled as positive or negative instances of a given class. To simplify
the explanation of the basic concepts, we will assume that all attributes are discrete.
Later, when the underlying principles become clear, we will slightly generalize the
approach to make it usable also in domains with continuous attributes or with mixed
sets of continuous and discrete attributes.
Decision Tree Figure 6.1 shows a few example decision trees that are capable of
dealing with the data from Table 6.1. The internal nodes represent attribute-value

DOI 10.1007/978-3-319-63913-0_6
114 6 Decision Trees
Table 6.1 Eight training crust filling

examples described by three
Example size shape size Class
symbolic attributes and
classified as positive and e1 big circle small pos
negative examples of a given e2 small circle small pos
class e3 big square small neg
e4 big triangle small neg
e5 big square big pos
e6 small square small neg
e7 small square big pos
e8 big circle big pos
tests, the edges indicate how to proceed in the case of diverse test results, and the
leafs1 contain class labels. An example to be classified is first subjected to the test
prescribed at the topmost node, the root. The result of this test then decides along
which edge the example is to be sent down, and the process continues until a leaf
node is reached. Once this happens, the example is labeled with the class associated
with this leaf.
Let us illustrate the process using the tree from Fig. 6.1b. The root asks about the
shape, each of whose three values is represented by one edge. In examples e1; e2,
and e8, we find the value shape=circle which corresponds to the left edge.
The example is sent down along this edge, ending in a leaf that labeled pos. This
indeed is the class common to all these three examples. In e4, shape=triangle,
and the corresponding edge ends in a leaf labeled neg—again, the correct class.
Somewhat more complicated is the situation with examples e3; e5; e6, and e7 where
shape=square. For this particular value, the edge ends not in a leaf, but only at a
test-containing node, this one inquiring about the value of filling-size. In the
case of e5 and e7, the value is big, which leads to a leaf labeled with pos. In the
other two examples, e3 and e6, the value is small, and this sends them to a leaf
labeled with neg.
We have shown that the decision tree from Fig. 6.1b identifies the correct class
for all training examples. By way of a little exercise, the reader is encouraged to
verify that the other trees shown in the picture are just as successful.2
Interpretability Comparing this classifier with those introduced earlier, we can see
one striking advantage: interpretability. If anybody asks why example e1 is deemed
positive, the answer is, “because its shape is circle.” Other classifiers do not
offer explanations of this kind. Especially the neural network is a real black box:
when presented with an example, it simply returns the class and never offers any
1
Both spellings are used: leaves and leafs. The latter is probably more appropriate because the
“leaf” in question is supposed to be a data abstraction that has nothing to do with the original
physical object.
2
In as sense, the decision tree can be seen as a simple mechanism for data compression.
6.1 Decision Trees as Classifiers 115
(A) (B)
crust
size
big small shape
circle triangle
sq
filling shape
size filling
big small
circle sq + size −
big small
filling
+ shape + size +
tri big small
−
circle
sq
+ − + −
(D)
filling
(C) big
size
small
filling
size crust
big small + size
small big
+ shape c
tri circle shape shape
sq
c tri
− + c sq
sq
− + + −
Fig. 6.1 Example decision trees for the “pies” domain. Note how they differ in size and in the
order of tests. Each of them classifies correctly all training examples listed in Table 6.1, tri, sq,
and c stand for triangle, square, and circle, respectively
insight as to why this particular label has been given preference over other labels.
The situation is not much better in the case of Bayesian and linear classifiers. Only
the k-NN classifier offers a semblance of a—rather rudimentary—argument. For
instance, one can say that, “x should be labeled with pos because this is the class
of the training example most similar to x.” Such a statement, however, is a far cry
from the explicit attribute-based explanation made possible by the decision tree.
One can go one step further and interpret a decision tree as a set of rules such
as “if shape=square and filling-size=big, then the example belongs
to class pos.” A domain expert inspecting these rules may then decide whether
they are intuitively appealing, and whether they agree with his or her “human
understanding” of the problem at hand. The expert may even be willing to suggest
improvements to the tree; for instance, by pointing out spurious tests that have found
their way into the data structure only on account of some random regularity in the
data.
Missing Edge The reader will recall that, in linear classifiers, an example may find
itself exactly on the class-separating hyperplane, in which case the class is selected
more or less randomly. Something similar occasionally happens in decision trees,
too. Suppose the tree from Fig. 6.1a is used to determine the class of the following
example:
.crust size D small/AND.shape D triangle/AND.filling size D small/
Let us follow the procedure step by step. The root inquires about crust-size.
Realizing that the value is small, the classifier sends the example down the right
edge, to the test on shape. Here, only two outcomes appear to be possible: circle
or square, but not triangle. The reason is, whoever created the tree had no idea
that an object with crust-size=small could be triangular: nothing of that kind
could be found in the training set. Therefore, there did not seem to be any reason for
creating the corresponding edge. And even if the edge were created, it would not be
clear where it should lead to.
The engineer implementing this classifier in a computer program must make sure
the program “knows” what to do in the case of “missing edges.” Choosing the class
randomly or preferring the most frequent class are the most obvious possibilities.
• Describe the mechanism that uses a decision tree to classify examples. Illustrate
the procedure using the decision trees from Fig. 6.1 and the training examples
from Table 6.1.
• What is meant by the statement that, “the decision tree’s choice of a concrete
class can be explained to the user”? Is something similar possible in the case of
the classifiers discussed in the previous chapters?
• Under what circumstances can a decision tree find itself unable to determine
an example’s class? How would you handle the situation if you were the
programmer?
6.2 Induction of Decision Trees 117
6.2 Induction of Decision Trees
We will begin with a very crude induction algorithm. Applying it to a training set, we
will realize that a great variety of alternative decision trees can be obtained. A brief
discussion will convince us that, among these, the smaller ones are to be preferred.
This observation will motivate an improved version of the technique, thus preparing
the soil for the following sections.
Divide and Conquer Let us try our hand at creating a decision tree manually.
Suppose we decide that the root node should test the value of shape. In the training
set, three different outcomes are found: circle, triangle, and square. For
each, the classifier will need a separate edge leading from the root. The first, defined
by shape=circle, will be followed by examples TC D fe1; e2; e8g; the second,
defined by shape=triangle, will be followed by TT D fe4g; and the last,
defined by shape=square, will be followed by TS D fe3; e5; e6; e7g. Each of
the three edges will begin at the root and end in another node, either an attribute test
or a leaf containing a class label.
Seeing that all examples in TC are positive, we will let this edge point at a
leaf labeled with pos. Similarly, the edge followed by the examples from TT will
point at a leaf labeled with neg. Certain difficulties will arise only in the case
of the last edge because TS is a mixture of both classes. To separate them, we
need another test, say, filling-size, to be placed at the end of the edge. This
attribute can acquire two values, small and big, dividing TS into two subsets. Of
these, TSS D fe3; e6g is characterized by filling-size=small; the other,
TSB D fe5; e7g, is characterized by filling-size=big. All examples in TSS
are positive, and all examples in TSB are negative. This allows us to let both edges
end in leafs, the former labeled with pos, the latter with neg. At this moment,
the tree-building process can stop because each training example presented to the
classifier thus created will reach a leaf.
The reader will have noticed that each node of the tree can be associated with
a set of examples that pass through it or end in it. Starting with the root, each test
divides the training set into disjoint subsets, and these into further subsets, and so
on until each subset is “pure” in the sense that all its examples belong to the same
class. This is why the approach is sometimes referred to as the divide-and-conquer
technique.
Alternative Trees In the process thus described, the (rather arbitrary) choice of
shape and filling-size resulted in the decision tree shown in Fig. 6.1b. To
get used to the mechanism, the student is encouraged to experiment with alternatives
such as placing at the root the test on crust-size or filling-size, and
considering different options for the tests at the lower level(s). Quite a few other
decision trees will thus be created—some of them depicted in Fig. 6.1.
That so many solutions can be found even in this very simple toy domain is a
food for thought. Is there a way to decide which trees are better? So, an improved
version of the divide-and-conquer technique should be able to arrive at a “good”
tree by design, and not by mere chance.
The Size of the Tree The smallest of the data structures in Fig. 6.1 consists of
two attribute tests; the largest, of five. Differences of this kind may have a strong
impact on the classifier’s behavior. Before proceeding to the various aspects of this
phenomenon, however, let us emphasize that the number of nodes in the tree is not
the only criterion of size. Just as important is the number of tests that have to be
carried out when classifying an average example.
For instance, in a domain where shape is almost always circle or
triangle (and only very rarely square), the average number of tests prescribed
by the tree from Fig. 6.1b will only slightly exceed 1 because both shape=circle
and shape=triangle immediately point at leafs with class labels. But if the
prevailing shape is square, the average number tests approaches 2. Quite often,
then, a bigger tree may result in fewer tests than a smaller one.
Small Trees Versus Big Trees There are several reasons why small decision trees
are preferred. One of them is interpretability. A human expert finds it easy to
analyze, explain, and perhaps even correct, a decision tree that consists of no more
than a few tests. The larger the tree, the more difficult this is.
Another advantage of small decision trees is their tendency to dispose of
irrelevant and redundant information. Whereas the relatively large tree from
Fig. 6.1a employs all three attributes, the smaller one from Fig. 6.1b is just as
good at classifying the training set—without ever considering crust-size. Such
economy will come handy in domains where certain attribute values are expensive
or time-consuming to obtain.
Finally, larger trees are prone to overfit the training examples. This is because the
divide-and-conquer method keeps splitting the training set into smaller and smaller
subsets, the number of these splits being equal to the number of attribute tests in the
tree. Ultimately, the resulting training subsets can become so small that the classes
may get separated by an attribute that only by chance—or noise—has a different
value in the remaining positive and negative examples.
Induction of Small Decision Trees When illustrating the behavior of the divide-
and-conquer technique on the manual tree-building procedure, we picked the
attributes at random. When doing so, we observed that some choices led to smaller
trees than others. Apparently, the attributes differ in how much information they
convey. For instance, shape is capable of immediately labeling some examples
as positive (if the value is circle) or negative (if the value is triangle); but
crust-size cannot do so unless assisted by some other attribute.
Assuming that there is a way to measure the amount of information provided by
each attribute (and such a mechanism indeed exists, see Sect. 6.3), we are ready to
formalize the technique for induction of decision trees by a pseudocode. The reader
will find it in Table 6.2.
6.3 How Much Information Does an Attribute Convey? 119
Table 6.2 Induction of decision trees

Let T be the training set.
grow(T):
(1) Find the attribute, at, that contributes the maximum information about the class
labels.
(2) Divide T into subsets, Ti , each characterized by a different value of at.
(3) For each Ti :
If all examples in Ti belong to the same class, then create a leaf labeled with
this class; otherwise, apply the same procedure recursively to each training subset:
grow(Ti ).
• Explain the principle of the divide-and-conquer technique for induction of
decision trees.
• What are the advantages of small decision trees in comparison to larger ones?
• What determines the size of the decision tree obtained by the divide-and-conquer
technique?
6.3 How Much Information Does an Attribute Convey?
To create a relatively small decision tree, the divide-and-conquer technique relies on

one critical component: the ability to decide how much information about the class
labels is conveyed by the individual attributes. This section introduces a mechanism
to calculate this quantity.
Information Contents of a Message Suppose we know that the training examples
are labeled as pos or neg, the relative frequencies of these two classes being
ppos and pneg , respectively.3 Let us select a random training example. How much
information is conveyed by the message, “this example’s class is pos”?
The answer depends on ppos . In the extreme case where all examples are known
to be positive, ppos D 1, the message does not tell us anything new. The reader
knows the example is positive even without being told so. The situation changes if
3
Recall that the relative frequency of pos is the percentage (in the training set) of examples labeled
with pos; this represents the probability that a randomly drawn example will be positive.
Table 6.3 Some values of ppos log2 ppos

the information contents
(measured in bits) of the 1.00 0 bits
message, “this randomly 0.50 1 bit
drawn example is positive.” 0.25 2 bits
0.125 3 bits
Note that the mes-
sage is impossible for
ppos D 0
both classes are known to be equally represented so that ppos D 0:5. Here, the guess
is no better than a flipped coin, so the message does offer some information. And
if a great majority of examples are known to be negative so that, say, ppos D 0:01,
then the reader is all but certain that the chosen example is going to be negative as
well; the message telling him that this is not the case is unexpected. And the lower
the value of ppos , the more information the message offers.
When quantifying the information contents of such a message, the following
formula has been found convenient:
Ipos D log2 ppos (6.1)
The negative sign compensates for the fact that the logarithm of ppos 2 .0; 1/ is
always negative. Table 6.3 shows the information contents for some typical values
of ppos . Note that the unit for the amount of information is 1 bit. Another comment:
the base of the logarithm being 2, it is fairly common to write log ppos instead of
the more meticulous log2 ppos .
Entropy (Average Information Contents) So much for the information contents
of a single message. Suppose, however, that the experiment is repeated many
times. Both messages will occur, “the example is positive,” and “the example is
negative,” the first with probability ppos , the second with probability pneg . The
average information contents of all these messages is then obtained by the following
formulas where the information contents of either message is weighted by its
probability (the T in the argument refers to the training set):
H.T/ D ppos log2 ppos pneg log2 pneg (6.2)
The attentive reader will protest that the logarithm of zero probability is not
defined, and Eq. (6.2) may thus be useless if ppos D 0 or pneg D 0. Fortunately,
a simple analysis (using limits and l’Hopital’s rule) will convince us that, for p
converging to zero, the expression p log p converges to zero, too, which means that
0 log 0 D 0.
H.T/ is called entropy of T. Its value reaches its maximum, H.T/ D 1, when
ppos D pneg D 0:5 (because 0:5 log 0:5 C 0:5 log 0:5 D 1); and it drops to its
minimum, H.T/ D 0, when either ppos D 1 or pneg D 1 (because 0 log 0 C 1
6.3 How Much Information Does an Attribute Convey? 121
log 1 D 0). By the way, the case with ppos D 1 or pneg D 1 is regarded as perfect
regularity because all examples belong to the same class; conversely, the case with
ppos D pneg D 0:5 is seen as a total lack of regularity.
Amount of Information Contributed by an Attribute The concept of entropy
(lack of regularity) will help us deal with the main question: how much does the
knowledge of the value of a discrete attribute, at, tell us about an example’s class?
Let us remind ourselves that at divides the training set, T, into subsets, Ti , each
characterized by a different value of at. Quite naturally, each subset will be marked
by its own probabilities (estimated by relative frequencies) of the two classes, pipos
and pineg . Based on the knowledge of these, Eq. (6.2) will give us the corresponding
entropies, H.Ti /.
Now let jTi j be the number of examples in Ti , and let jTj be the number of
examples in the whole training set, T. The probability that a randomly drawn
training example will be in Ti is estimated as follows:
jTi j
Pi D (6.3)
jTj
We are ready to calculate the weighted average of the entropies of the subsets.
H.T; at/ D †i Pi H.Ti / (6.4)
The obtained result, H.T; at/, is the entropy of a system where not only the class
labels, but also the values of at are known for each training example. The amount
of information contributed by at is then the difference between the entropy before
at has been considered and the entropy after this attribute has been considered:
I.T; at/ D H.T/ H.T; at/ (6.5)
It would be easy to prove that this difference cannot be negative; information

can only be gained, never lost, by considering at. In certain rare cases, however,
I.T; at/ D 0, which means that no information has been gained, either.
Applying Eq. (6.5) separately to each attribute, we can find out which of them
provides the maximum amount of information, and as such is the best choice for the
“root” test in the first step of the algorithm from Table 6.2.
The procedure just described is summarized by the pseudocode in Table 6.4.
The process starts by the calculation of the entropy of the system where only
class percentages are known. Next, the algorithm calculates the information gain
conveyed by each attribute. The attribute that offers the highest information gain is
deemed best.
Illustration Table 6.5 shows how to use the mechanism for the selection of the
most informative attribute in the domain from Table 6.1. At the beginning, the
entropy, H.T/, of the system without attributes is established. Then, we observe
that, for instance, the attribute shape divides the training set into three subsets. The
Table 6.4 The algorithm to find the most informational attribute
1. Calculate the entropy of the training set, T, using the percentages, ppos and pneg , of
the positive and negative examples:
H.T/ D ppos log2 ppos pneg log2 pneg
2. For each attribute, at, that divides T into subsets, Ti , with relative sizes Pi , do the
following:
(i) calculate the entropy of each subset, Ti ;
(ii) calculate the average entropy: H.T; at/ D †i Pi H.Ti /;
(iii) calculate information gain: I.T; at/ D H.T/ H.T; at/
3. Choose the attribute with the highest value of information gain.
average of their entropies, H.T; shape/, is calculated, and the difference between
H.T/ and H.T; shape/ gives the amount of information conveyed by this attribute.
Repeating the procedure for crust-size and filling-size, and comparing
the results, we realize the shape contributes more information than the other two
attributes, and this is why we choose shape for the root test.
This, by the way, is how the decision tree from Fig. 6.1b was obtained.
• What do we mean when we talk about the “amount of information conveyed by
a message”? How is this amount determined, and what units are used?
• What is entropy and how does it relate to the frequency of the positive and
negative examples in the training set?
• How do we use entropy when assessing the amount of information contributed
by an attribute?
6.4 Binary Split of a Numeric Attribute
The entropy-based mechanism from the previous section requested that all attributes
should be discrete. With a little modification, however, the same approach can be
applied to continuous attributes as well. All we need is to convert them to boolean
attributes.
6.4 Binary Split of a Numeric Attribute 123
Table 6.5 Illustration of the search for the most informative attribute
Example crust shape filling Class

size size
e1 big circle small pos
e2 small circle small pos
e3 big square small neg
e4 big triangle small neg
e5 big square big pos
e6 small square small neg
e7 small square big pos
e8 big circle big pos
Here is the entropy of the training set where only class labels are known:
H.T/ D ppos log2 ppos pneg log2 pneg

D .5=8/ log.5=8/ .3=8/ log.3=8/ D 0:954
Next, we calculate the entropies of the subsets defined by the values of shape:
H.shape=square/ D .2=4/ log.2=4/ .2=4/ log.2=4/ D 1

H.shape=circle/ D .3=3/ log.3=3/ .0=3/ log.0=3/ D 0
H.shape=triangle/ D .0=1/ log.0=1/ .1=1/ log.1=1/ D 0
From these, we obtain the average entropy of the system where the class labels and the
value of shape is known:
H.T; shape/ D .4=8/ 1 C .3=8/ 0 C .1=8/ 0 D 0:5
Repeating the same procedure for the other two attributes, we obtain the following:
H.T; crust size/ D 0:951

H.T; filling size/ D 0:607
These values give the following information gains:
I.T; shape/ D H.T/ H.T; shape/ D 0:954 0:5 D 0:454

I.T; crust size/ D H.T/ H.T; crust size/ D 0:954 0:951 D 0:003
I.T; filling size/ D H.T/ H.T; filling size/ D 0:954 0:607 D 0:347
We conclude that maximum information is contributed by shape.

Converting a Continuous Attribute to a Boolean One Let us denote the contin-

uous attribute by x. The trick is to choose a threshold, , and then decide that if
x < , then the value of the newly created boolean attribute is true, and otherwise it
is false (or vice versa).
Simple enough. But what concrete to choose? Surely there are many of them?
Here is one possibility.
Suppose that x has a different value in each of the N training examples. Let us
sort these values in ascending order, denoting by x1 the smallest, and by xN the
highest. Any pair of neighboring values, xi and xiC1 , then defines a threshold, i D
.xi C xiC1 /=2. For instance, a four-example training set where x has values 3, 4,
6, and 9 leads us to consider 1 D 3:5; 2 D 5:0, and 3 D 7:5. For each of these
N 1 thresholds, we calculate the amount of information contributed by the boolean
attribute thus defined, and then choose the threshold where the information gain is
maximized.
Candidate Thresholds The approach just described deserves to be criticized for its
high computational costs. Indeed, in a domain with a hundred thousand examples
described by a hundred attributes (nothing extraordinary), the information contents
of 105 102 D 107 different thresholds would have to be calculated. Fortunately,
mathematicians have been able to prove that a great majority of these thresholds can
just as well be ignored. This reduces the costs to a mere fraction.
The principle is illustrated in Table 6.6. In the upper part, 13 values of x are
ordered from left to right, each labeled with the class (positive or negative) of
the training example in which the value was found. And here is the rule: the best
threshold never finds itself between values that are labeled with the same class. This
means that it is enough to investigate the contributed information only for locations
between values with opposite class labels. In the specific case shown in Table 6.6,
only three candidate thresholds, 1 ; 2 , and 3 , need to be investigated (among the
three, 1 is shown to be best).
The Root of a Numeric Decision Tree The algorithm summarized by the pseu-
docode in Table 6.7 determines the best attribute test for the root of a decision tree
in a domain where all attributes are continuous. Note that the test consists of a pair,
[ati , ij ], where ati is the selected attribute and ij is the best threshold found for this
attribute. If an example’s value of the i-th attribute is below the threshold, ati < ij ,
the left branch of the decision tree is followed; otherwise, the right branch is chosen.
• Explain why this section suggested to divide the domain of a continuous attribute
into two subdomains.
• What mathematical finding has reduced the number of thresholds that need to be
investigated?
6.4 Binary Split of a Numeric Attribute 125
Table 6.6 Illustration of the search for the best threshold

The values of attribute x are sorted in ascending order. The candidate thresholds are
located between values labeled with opposite class labels.
+ ++ + + ___ __ + + _
x
θ1 θ2 θ3
Here is the entropy of the training set, ignoring attribute values:
H.T/ D pC log pC p log p

D .7=13/ log.7=13/ .6=13/ log.6=13/ D 0:9957
Here are the entropies of the training subsets defined by the three candidate thresholds:
H.x < 1 / D .5=5/ log.5=5/ .0=5/ log.0=5/ D 0

H.x > 1 / D .2=8/ log.2=8/ .6=8/ log.6=8/ D 0:8113
H.x < 2 / D .5=10/ log.5=10/ .5=10/ log.5=10/ D 1

H.x > 2 / D .2=3/ log.2=3/ .1=3/ log.1=3/ D 0:9183
H.x < 3 / D .7=12/ log.7=12/ .5=12/ log.5=12/ D 0:9799

H.x > 3 / D .0=1/ log.0=1/ .1=1/ log.1=1/ D 0
Average entropies associated with the individual thresholds:
H.T; 1 / D .5=13/ 0 C .8=13/ 0:8113 D 0:4993

H.T; 2 / D .10=13/ 1 C .3=13/ 0:9183 D 0:9811
H.T; 3 / D .12=13/ 0:9799 C .1=13/ 0 D 0:9045
Information gains entailed by the individual candidate thresholds:
I.T; 1 / D H.T/ H.T; 1 / D 0:9957 0:4993 D 0:4964

I.T; 2 / D H.T/ H.T; 2 / D 0:9957 0:9811 D 0:0146
I.T; 3 / D H.T/ H.T; 3 / D 0:9957 0:9045 D 0:0912
Threshold 1 gives the highest information gain.

Table 6.7 Algorithm to find the best numeric-attribute test

1. For each attribute ati :
(i) Sort the training examples by the values of ati ;
(ii) Determine the candidate thresholds, ij , as those lying between examples with
opposite labels;
(iii) For each ij , determine the amount of information contributed by the boolean
attribute thus created.
2. Choose the pair Œati ; ij that offers the highest information gain.
6.5 Pruning
Section 6.2 extolled the virtues of small decision trees: interpretability, removal of
irrelevant and redundant attributes, reduced danger of overfitting. These were the
arguments that motivated the use of information theory in the course of decision-
tree induction; they also motivate the step that follows: pruning.
Fig. 6.2 A simple approach t1

to pruning will replace a 1 0
subtree with a leaf
t2 t5
1 0 1 0 1
-
+
+
t3 t6
1 0 1 0
t4 - -
+
1 0
-
+
The Essence Figure 6.2 will help us explain the principle. On the left is the original
decision tree whose six attribute tests are named t1 ; : : : t6 . On the right is a pruned
version. Note that the subtree rooted in test t3 in the original tree is in the pruned tree
replaced with a leaf labeled with the negative class; and the subtree rooted in test t6
is replaced with a leaf labeled with the positive class. The reader can see the point:
pruning consists of replacing one or more subtrees with leafs, each labeled with
the class most common among the training examples that reach—in the original
classifier—the removed subtree.
This last idea sounds counterintuitive: the induction mechanism seeks to create
a decision tree that scores zero errors on the training examples, but this perfection
may be lost in the pruned tree! But the practically minded engineer is not alarmed.
The ultimate goal is not to classify the training examples (their classes are
known anyway). Rather, we want a tool capable of labeling future examples. And
experience shows that this kind of performance is often improved by reasonable
pruning.
6.5 Pruning 127
Error Estimate Pruning is typically carried out in a sequence of steps: first replace
with a leaf one subtree, then another, and so on, as long as the replacements appear
to be beneficial according to some reasonable criterion. The term “beneficial” is
meant to warn us that small-tree advantages should not be outbalanced by reduced
classification performance.
Which brings us to the issue of error estimate. The principle is illustrated in
Fig. 6.2. Let m be the number of training examples that reach test t3 in the decision
tree on the left. If we replace the subtree rooted in t3 by a leaf (as happened in the
tree on the right), some of these m examples may become misclassified. Denoting
the number of these misclassified examples by e, we may be tempted to estimate
the probability that an example will be misclassified at this leaf by the relative
frequency: e=m. But admitting that small values of m may render this estimate
problematic, we prefer the following formula where N is the total number of training
examples:
eC1
Eestimate D (6.6)
NCm
The attentive reader may want to recall (or re-read) what Sect. 2.3 had to say
about the difficulties of probability estimates of rare events.
Error Estimates for the Whole Tree Once again, let us return to Fig. 6.2. The tree
on the left has two subtrees, one rooted at t2 , the other at t5 . Let m2 and m5 be the
numbers of the training examples reaching t2 and t5 , respectively; and let E2 and
E5 be the error estimates of the two subtrees, obtained by Eq. (6.6). For the total of
N D m2 C m5 training examples, the error rate of the whole subtree is estimated as
the weighted average of the two subtrees:
m2 m5
ER D E2 C E5 (6.7)
N N
Of course, in a situation with more than just two subtrees, the weighted average
has to be taken over all of them. This should present no major difficulties.
As for the values of E2 and E5 , these are obtained from the error rates of the
specific subtrees, and these again from the error rates of their sub-subtrees, and so
on, all the way down to the lowest level tests. The error-estimating procedure is a
recursive undertaking.
Suppose that the tree to be pruned is the one rooted at t3 , which happens to be one
of the two children of test t2 . The error estimate for t2 is calculated as the weighted
average of E3 and the error estimate for the other child of t2 (the leaf labeled with
the positive class). The resulting estimate would then be combined with E5 as shown
above.
Post-pruning The term refers to the circumstance that the decision tree is pruned
after it has been fully induced from data (an alternative is the subject of the next
subsection). We already know that the essence is to replace a subtree with a leaf
labeled with the class most frequent among the training examples reaching that leaf.
Since there are usually several (or many) subtrees that can thus be replaced, a choice
has to be made; and the existence of a choice means we need a criterion to guide
our decision.
Here is one possibility. We know that pruning is likely to change the classifier’s
performance. One way to assess this change is to compare the error estimate of the
decision tree after the pruning with that of the tree before the pruning:
D D Eafter Ebefore (6.8)
From the available pruning alternatives, we choose the one where this difference
is the smallest, Dmin ; but we carry out the pruning only if Dmin < c, where c is a user-
set threshold for how much performance degradation can be tolerated in exchange
for the tree’s compactness. The mechanism is then repeated, with the decision tree
becoming smaller and smaller, the stopping criterion being imposed by the constant
c. Thus in Fig. 6.2, the first pruning step might have removed the subtree rooted at t3 ;
and the second step, the subtree rooted at t6 . Here the procedure was stopped because
any further attempt at pruning resulted in a tree whose error estimate increased too
much: the difference between the estimated error of the final (pruned) tree and that
of the original tree on the left of Fig. 6.2 exceeded the user’s threshold: D > c.
The principle is summarized by the pseudocode in Table 6.8.
On-line Pruning In the divide-and-conquer procedure, each subsequent attribute
divides the set of training examples into smaller and smaller subsets. Inevitably,
the evidence supporting the choice of the tests at lower tree-levels will be weak.
When a tree node is reached by only, say, two training examples, one positive
and one negative, a totally irrelevant attribute may by mere coincidence succeed in
separating the positive example from the negative. The only “benefit” to be gained
from adding this test to the decision tree is training-set overfitting.
The motivation behind on-line pruning is to make sure this situation is prevented.
Here is the rule: if the training subset is smaller than a user-set minimum, m, stop
further expansion of the tree.
Impact of Pruning How far the pruning goes is controlled by two parameters: c in
post-pruning, and m in on-line pruning. In both cases, higher values result in smaller
trees.
Table 6.8 The algorithm for decision-tree pruning

c : : : a user-set constant
(1) Estimate the error rate of the original decision tree. Let its value be denoted by Ebefore .
(2) Estimate the error rates of the trees obtained by alternative ways of pruning the
original tree.
(3) Choose the pruning after which the estimated error rate experiences minimum
increase, Dmin D Ebefore Eafter , but only if Dmin < c.
(4) Repeat steps (2) and (3) as long as Ebefore Eafter < c.
6.5 Pruning 129
The main reason why pruning tends to improve classification performance on

future examples is that the removal of low-level tests, which have poor statistical
support, usually reduces the danger of overfitting. This, however, works only up to
a certain point. If overdone, a very high extent of pruning can (in the extreme) result
in the decision being replaced with a single leaf labeled with the majority class.
Such classifier is unlikely to be useful.
Figure 6.3 shows the effect that gradually increased pruning typically has on
classification performance. Along the horizontal axis is plotted the extent of pruning
as controlled by c or m or both. The vertical axis represents the error rate measured
on the training set as well as on some hypothetical testing set (the latter consisting
of examples that have not been used for learning, but whose class labels are known).
On the training-set curve, error rate is minimized when there is no pruning
at all. More interesting, however, is the testing-set curve. Its shape is telling us
that the unpruned tree usually scores poorly on testing data, which is explained
by the unpruned tree’s tendency to overfit the training set, a phenomenon that
can be reduced by increasing the extent of pruning. Excessive pruning, however,
will remove attribute tests that do carry useful information, and this will have a
detrimental effect on classification performance.
By the way, the two curves can tell us a lot about the underlying data. In some
applications, even very modest pruning will impair error rate on testing data; for
instance, in a noise-free domain with a relatively small training set.
Another thing to notice is that the error rate on the testing set is almost always
greater than the error rate on the training set.
• What are the potential benefits of decision-tree pruning?
• How can we estimate the tree’s error rate on future data? Write down the formula
and explain how it is used.
Fig. 6.3 With the growing error rate

extent of pruning, error rate
on the testing set usually
drops, then starts growing
testing set
again. Error rate on the
training set usually increases training set
monotonically
extent of pruning
• Describe the principle of post-pruning and the principle of on-line pruning.

• What parameters control the extent of pruning? How do they affect the error rate
on the training set, and how do they affect the error rate on the testing set?
6.6 Converting the Decision Tree into Rules
One of the advantages of decision trees in comparison with the other classifiers is
their interpretability. Any sequence of tests along the path from the root to a leaf
represents an if-then rule, and this rule explains why the classifier has labeled a
given example with this or that class.
Rules Generated by a Decision Tree The reader will find it easy to convert a
decision tree to a set of rules. It is enough to notice that a leaf is reached through a
series of edges whose specific choice is determined by the results of the attribute
tests encountered along the way. Each leaf is thus associated with a concrete
conjunction of test results.
For the sake of illustration, let us write down the complete set of rules for the
pos class as obtained from the decision tree in Fig. 6.1a.
if crust-size=big AND filling-size=big then pos
if crust-size=big AND filling-size=small AND shape=circle
then pos
if crust-size=small AND shape=circle then pos
if crust-size=small AND (shape=square OR triangle)
AND filling-size=big then pos
else neg
Note the default class, neg, in the last line. An example is labeled with the
default class if all rules fail, which means that the value of the if -part of each rule
is false. We notice that, in this two-class domain, we need to consider only the
rules resulting in the pos class, the other class being the default option. We could
have done it the other way round, considering only the rules for the neg class, and
making pos the default class. This would actually be more economical because
there are only two leafs labeled with the neg, and therefore only two corresponding
rules. The reader is encouraged to write down these two rules by way of a simple
exercise.
At any rate, the lesson is clear: in a domain with K classes, only the rules for
K 1 classes are needed, the last class serving as the default.
Pruning the Rules The tree post-pruning mechanism described earlier replaced
a subtree with a leaf. This means that lower-level tests were the first to go, the
technique being unable to remove a higher-level node before the removal of the
nodes below it. The situation is similar in on-line pruning.
Once the tree has been converted to rules, however, pruning gains in flexibility:
any test in the if -part of any rule is a potential candidate for removal; and entire
6.6 Converting the Decision Tree into Rules 131
Table 6.9 The algorithm for rule pruning

Re-write the decision tree as a set of rules.
Let c be a user-set constant controlling the extent of pruning
(1) In each rule, calculate the increase in error estimate brought about by the removal of
individual tests.
(2) Choose those removals where this increase, Dmin is smallest. Remove the tests, but
only if Dmin < c.
(3) In the set of rules, search for the weakest rules to be removed.
(4) Choose the default class.
(5) Order the rules according to their strengths (how many training examples they cover).
Table 6.10 Illustration of the algorithm for rule pruning

Suppose that the decision from the left part of Fig. 6.2 has been converted into the
following set of rules (neg is the default label to be used when the if-parts of all rules are
false).
t1 ^ t2 ! pos
t1 ^ :t2 ^ t3 ^ t4 ! pos
:t1 ^ t5 ^ t6 ! pos
else neg
Suppose that the evaluation of the tests in the rules has resulted in the conclusion that t3
in the second rule and t5 in the third rule can be removed without a major increase in the
error estimate. We obtain the following set of modified rules.
t1 ^ t2 ! pos
t1 ^ :t2 ^ t4 ! pos
:t1 ^ t6 ! pos
else neg
The next step can reveal that the second (already modified) rule can be removed without
a major increase in the error estimate. After the removal, the set of rules will look as
follows.
t1 ^ t2 ! pos
:t1 ^ t6 ! pos
else neg
This completes the pruning.
rules can be deleted. This is done by the rule-pruning algorithm summarized by

the pseudocode in Table 6.9 and illustrated by the example in Table 6.10. Here,
the initial set of rules was obtained from the (now familiar) tree in the left part of
Fig. 6.2. The first pruning step removes those tests that do not appear to contribute
much to the overall classification performance; the next step deletes the weakest
rules.
We haste to admit, however, that the price for this added flexibility is a significant
increase in computational costs.
• Explain the mechanism that converts a decision tree to a set of rules. How many
rules are thus obtained? What is the motivation behind such conversion?
• What is meant by the term, default class? How would you choose it?
• Discuss the possibilities of rule-pruning. In what sense can we claim that rule-
pruning offers more flexibility than decision-tree pruning? What is the price for
this increased flexibility?
• In decision trees, the attribute values are tested one at a time, the result of each
test indicating what should happen next: either another attribute test, or a decision
about the class label if a leaf has been reached. One can say that a decision tree
consists of a set of partially ordered set of tests, each sequence of tests defining
one branch in the tree terminated by a leaf.
• From a typical training set, many alternative decision trees can be created.
As a rule, smaller trees are to be preferred, their main advantages being
interpretability, removal of irrelevant and redundant attributes, and lower danger
of overfitting noisy training data.
• The most typical procedure for induction of decision trees from data proceeds in a
recursive manner, always seeking to identify the attribute that conveys maximum
information about the class label. This approach tends to make the induced
decision trees smaller. The “best” attribute is identified by simple formulas
borrowed from information theory.
• An important aspect of decision-tree induction is pruning. The main motivation
is to make sure that all tree branches are supported by sufficient evidence. Further
on, pruning reduces the tree size which has certain advantages (see above). Two
generic types of pruning exist. (1) In post-pruning, the tree is first fully developed,
and then pruned. (2) In on-line pruning (which is perhaps a bit of a misnomer),
the development of the tree is stopped once the training subsets used to determine
the next attribute test become too small. In both cases, the extent of pruning is
controlled by user-set parameters (denoted c and m, respectively).
• A decision tree can be converted to a set of rules that can further be pruned.
In a domain with K classes, it is enough to specify the rules for K 1 classes,
the remaining class becoming the default class. The rules are usually easier to
interpret. Rule-pruning algorithms sometimes lead to more compact classifiers,
though at significantly increased computational costs.
Historical Remarks The idea behind decision trees was first put forward by
Hoveland and Hund in the late 1950s. The work was later summarized in the
book Hunt et al. [39] that reports experience with several implementations of
their Concept Learning System (CLS). Friedman et al. [30] developed a similar
approach independently. An early high point of the research was reached by
Breiman et al. [11] where the system CART is described. The idea was then
imported to the machine-learning world by Quinlan [75, 76]. Perhaps the most
famous implementation is C4.5 from Quinlan [78]. This chapter was based on a
simplified version of C4.5.
Fig. 6.4 Another example

with wooden and plastic
wood plastic
circles
Exercises
1. In Fig. 6.4, eight training examples are described by two attributes, size and
color, the class label being the material: either wood or plastic.
• What is the entropy of the training set when only the class labels are
considered (ignoring the attribute values)?
• Using the information-based mechanism from Sect. 6.3, decide which of the
two attributes better predicts the class
2. Take the decision tree from Fig. 6.1a and remove from it the bottom-right test
on filling-size. Based on the training set from Table 6.1, what will be the
error estimate before and after this “pruning”?
3. Choose one of the decision trees in Fig. 6.1 and convert it to a set of rules. Pick
one of these rules and decide which of its tests can be removed with the minimum
increase in the estimated error.
4. Consider a set of ten training examples. Suppose there is a continuous attribute
that has the following values: 3:6; 3:2; 1:2; 4:0; 0:8; 1:2; 2:8; 2:4; 2; 2; 1:0. Sup-
pose that the first five of these examples, and also the last one, are positive, all
other examples being negative. What will be the best binary split of the range of
this attribute’s values?
1. The baseline performance criteria used for the evaluation of decision trees are
error rate and the size of the tree (the number of nodes). These, however, may
not be appropriate in certain domains. Suggest applications where either the size
of the decision tree or its error rate may be less important. Hint: Consider the
costs of erroneous decisions and the costs of obtaining attribute values.
2. What are likely to be the characteristics of a domain where a decision tree clearly
outperforms the baseline 1-NN classifier? Hint: Consider such characteristics as
noise, irrelevant attributes, or the size of the training set; and then make your own
judgement as to what influence each of them is likely to have on the classifier’s
behavior.
3. On what kind of data may a linear classifier do better than a decision tree? Give at
least two features characterizing such data. Rely on the same hint as the previous
question.
4. Having found the answers to the previous two questions, you should be able
to draw the logical conclusion: applying to the given data both decision-tree
induction and linear-classifier induction, what will their performance betray
about the characteristics of the data?
5. The decision tree as described in this chapter gives only “crisp” yes-or-no
decisions about the given example’s class (in this sense, one can argue that
Bayesian classifiers or multilayer perceptrons are more flexible). By way of
mitigating this weakness, suggest a mechanism that would modify the decision-
trees framework so as to give, for each example, not only the class label, but also
the classifier’s confidence in this class label.
1. Implement the baseline algorithm for the induction of decision trees and test
its behavior on a few selected domains from the UCI repository.4 Compare the
results with those achieved by the k-NN classifier.
2. Implement the simple pruning mechanism described in this chapter. Choose a
data file from the UCI repository. Run several experiments and observe how
different extent of pruning affects the error rate on the training and testing sets.
3. Choose a sufficiently large domain from the UCI repository. Put aside 30% of
the examples for testing. For training, use 10%, 20%, . . . 70% of the remaining
examples, respectively. Plot a graph where the horizontal axis gives the number
of examples, and the vertical axis gives the computational time spent on the
induction. Plot another graph where the vertical axis will give the error rate on
the testing set. Discuss the obtained results.
4
Chapter 7
Computational Learning Theory
As they say, nothing is more practical than a good theory. And indeed, mathematical
models of learnability have helped improve our understanding of what it takes to
induce a useful classifier from data, and, conversely, why the outcome of a machine-
learning undertaking so often disappoints. And so, even though this textbook does
not want to be mathematical, it cannot help introducing at least the basic concepts
of the computational learning theory.
At the core of this theory is the idea of PAC learning, a paradigm that makes
it possible to quantify learnability. Restricting itself to domains with discrete
attributes, the first section of this chapter derives a simple expression that captures
the mutual relation between the training-set size and the induced classifier’s error
rate. Some consequences of this formula are briefly discussed in the two sections
that follow. For domains with continuous attributes, the so-called VC-dimension is
then introduced.
7.1 PAC Learning
Perhaps the most useful idea contributed by computational learning theory is that of
“probably approximate learning,” sometimes abbreviated as PAC learning. Let us
first explain the underlying principles and derive a formula that will then provide
some useful guidance.
Assumptions and Definitions The analysis will be easier if we build it around
a few simplifying assumptions. First, the training examples—as well as all future
examples—are completely noise-free. Second, all attributes are discrete (none of
them is continuous-valued). Third, the classifier acquires the formof a logical

DOI 10.1007/978-3-319-63913-0_7
138 7 Computational Learning Theory
expression of attribute values; the expression is true for positive examples and false
for negative examples. And, finally, there exists at least one expression that correctly
classifies all training examples.1
Each of the logical expressions can then be regarded as one hypothesis about
what the classifier should look like. Together, all of the hypotheses form a hypothesis
space whose size (the number of distinct hypotheses) is jHj. Under the assumptions
listed above, jHj is a finite number.
Inaccurate Classifier May Still Succeed on the Training Set Available training
data rarely exhaust all subtleties of the underlying class; a classifier that labels
correctly all training examples may still perform poorly in the future. The frequency
of these mistakes is usually lower if we add more training examples because these
additions may reflect aspects that the original data failed to represent. The rule of
thumb is actually quite simple: the more examples we have, the better the induced
classifier.
How many training examples will give us a fair chance of future success? To
find the answer, we will first consider a hypothetical classifier whose error rate
on the entire instance space is greater than some predefined
. Put another way.
the probability that this classifier will label correctly a randomly picked example
is less than 1
. Taking this reasoning one step further, the probability, P, that
this imperfect classifier will label correctly m random examples is bounded by the
following expression:
P .1
/m (7.1)
And here is what this means: with probability P .1

/m , an entire training
set consisting of m examples will be correctly classified by a classifier whose error
rate actually exceeds
. Of course, this probability is for a realistic m very low. For
instance, if
D 0:1 and m D 20 (which is a very small set), we will have P < 0:122.
If we increase the training-set size to m D 100 (while keeping the error rate bound
by
D 0:1), then P drops to less than 104 . This is indeed small; but low probability
is not the same as impossibility.
Eliminating Poor Classifiers Suppose that an error rate greater than
is deemed
unacceptable. What are the chances that a classifier with performance this poor is
induced from the given training set, meaning that it classifies correctly all training
examples?
The hypothesis space consists of jHj classifiers. Let us consider the theoretical
possibility that we evaluate all these classifiers on the m training examples, and
then keep only those classifiers that have never made any mistake. Amongthese
1
The reader will have noticed that all these requirements are satisfied by the “pies” domain from
Chap. 1.
7.1 PAC Learning 139
“survivors,” some will disappoint in the sense that, while being error-free on the
training set, their error rates on the entire instance space actually exceed
. Let there
be k such classifiers.
The concrete value of k cannot be established without evaluating each single
classifier on the entire instance space. This being impossible, all we can say is that
k jHj, which is a somewhat better situation because jHj is known in many realistic
cases, or at least can be calculated.2
Let us re-write the upper bound on the probability that at least one of the k
offending classifiers will be error-free on the m training examples.
P k.1
/m jHj.1
/m (7.2)
With this, we have established an upper bound on the probability that m training
examples will succeed in eliminating all classifiers whose error rate exceeds
.
To become useful, the last expression has to be modified. We know from
mathematics that 1
< e
, which means that (1
/m < em
. With this in
mind, we are able to express the upper bound in an exponential form:
P jHj em
(7.3)
Suppose we want this probability to be lower than some user-set ı:
jHj em
ı (7.4)
Taking the logarithm of both sides, and rearranging the terms, we obtain the
formula that we will work with in the next few pages:
1 1
m> .ln jHj C ln / (7.5)

ı
“Probably Approximately Correct” Learning The main reason we have gone
through all the details of the derivation is that the reader thus gets a better grasp of
the meanings and interpretations of the involved variables which may otherwise be
a bit confusing. This is also why these variables are summarized in Table 7.1 for
quick reference.
We have now reached a stage where we are able to define some important
concepts. A classifier with error rate below
is deemed approximately correct; and
ı is the probability that this approximately correct classifier will be induced from m
training examples (m being a finite number). Hence the name of the whole paradigm:
probably approximately correct learning, or simply PAC learning. For the needs of
this chapter, we will say that a class is not PAC-learnable if the number of examples
2
Recall that in the “pies” domain from Chap. 1, the size of the hypothesis space was jHj D 2108 .
Of these hypotheses, 296 classified correctly the entire training set.
Table 7.1 The variables m ::: The number of the training examples
involved in our studies of
jHj ::: The size of the hypothesis space
PAC-learnability

::: The classifier’s maximum permitted error rate
ı ::: The probability that a classifier with error rate
greater than
is error-free on the training set
needed to satisfy the given (

; ı)-requirements is so high that we cannot expect a
training set of this size ever to be available—or, if it is available, then the learning
software will need impracticably long time (say, thousands of years) to induce from
it the classifier.
Interpretation Inequality (7.5) specifies how many training examples, m, are
needed if, with probability at least ı, a classifier with error rate lower than
is to
be induced. Note that this result does not depend on the concrete machine-learning
technique, only on the size, jHj, of the hypothesis space defined by the given type
of classifier.
An important thing to remember is that m grows linearly in 1=
. For instance,
if we strengthen the limit on the error rate from
D 0:2 to
D 0:1, we will need
(at least in theory) twice as many training examples to have the same chance, ı, of
success. At the same time, the reader should also notice that m is less sensitive to
changes in ı, growing only logarithmically in 1=ı.
However, we must not forget that our derivation was something like a worst-case
analysis. As a result, the bound it has given us is less tight than a perfectionist might
desire. For instance, the derivation allowed the possibility that k D jHj, which is
clearly too pessimistic for the vast majority of practical applications. Inequality (7.5)
should therefore never be interpreted as telling the engineer how many training
examples to use. Rather, it should be seen as a guideline that makes it possible
to compare the learnability of alternative classifier types. We will pursue this idea in
the next section.
• What is the meaning of the four variables (m; jHj;
; and ı) used in Inequal-
ity (7.5)?
• What does the term PAC learning refer to? Under what circumstances do we say
that a class is not PAC-learnable?
• Derive Inequality (7.5). Discuss the meaning and practical benefits of this result,
and explain why it should only be regarded as a worst-case analysis.
7.2 Examples of PAC Learnability 141
7.2 Examples of PAC Learnability
Inequality (7.5) tells us how learnability, defined by the (

; ı)-requirements, depends
on the size of the hypothesis space. Let us illustrate this result on two concrete types
of classifiers.
Conjunctions of Boolean Attributes: Hypothesis Space Suppose that all
attributes are boolean, that all data are noise-free, and that an example’s class
is known to be determined by a logical conjunction of attribute values: if true, then
the example is positive, otherwise it is deemed negative. For instance, the labels of
the training examples may be determined by the following expression:
att1 = true AND att3 = false
This means that an example is labeled as pos if the value of the first attribute is
true and the value of the third attribute is false, regardless of the value of any other
attribute. An example that fails to satisfy these two conditions is labeled as neg.
The task for the machine-learning program is to find a conjunction that correctly
labels all training examples. The set of all conjunctions permitted by our informal
definition forms the hypothesis space, jHj. What is the size of this space?
In a logical conjunction of the kind specified above, each attribute is either true
or false or irrelevant. This gives us three possibilities for the first attribute, times
three possibilities for the second, and so on, times three possibilities for the last,
n-th, attribute. The size of the hypothesis space is therefore jHj D 3n .
Conjunctions of Boolean Attributes: PAC-Learnability Suppose that a training
set consisting of noise-free examples is presented to some machine-learning pro-
gram that is capable of inducing classifiers of the just-defined form.3 To satisfy the
last of the assumptions listed at the beginning of Sect. 7.1, we will assume that at
least one of the logical conjunctions classifies correctly all training examples.
Since ln jHj D ln 3n D n ln 3, we can re-write Inequality (7.5) in the following
form:
1 1
m> .n ln 3 C ln / (7.6)

ı
With this, we have obtained a conservative lower bound on the number of
training examples that are needed if our (
; ı)-requirements are to be satisfied: with
probability ı, the induced classifier (error-free on the training set) will exhibit error
rate less than
on the entire instance space. Note that the value of this expression
grows linearly in the number of attributes, n. Theoretically speaking, then, if n is
doubled, then twice as many training examples will be needed if, with probability
limited by ı, classifiers with error rates above the predefined
are to be weeded out.
3
For instance, a variation of the hill-climbing search from Sect. 1.2 might be used to this end.
Any Boolean Function: Hypothesis Space Let us now investigate a broader

class of classifiers, namely those defined by any boolean function, allowing for all
three basic logical operators (AND, OR, and NOT) as well as any combination of
parentheses. As before, we will assume that the examples are described by n boolean
attributes, and that they are noise-free.
What is the size of this hypothesis space?
From n boolean attributes, 2n different examples can be created. This defines the
size of the instance space. For any subset of these 2n examples, there exists at least
one logical function that is true for all examples from this subset (labeling them as
pos) and false for all examples from outside this subset (labeling them as neg).
Two logical functions will be regarded as identical from the classification point of
view if each of them labels any example with the same class; that is, if they never
differ in their “opinion” about any example’s class. The number of logical functions
that are mutually distinct in their classification behavior then has to be the same as
the number of the subsets of the instance space.
A set consisting of X elements is known to have 2X subsets. Since our specific
instance space consists of 2n examples, the number of its subsets is 22 —and this is
n
the size of our hypothesis space:
jHj D 22
n
(7.7)
Any Boolean Function: PAC-Learnability Since ln jHj D ln 22

n
D 2n ln 2,
Inequality (7.5) acquires the following form:
1 n 1
m> .2 ln 2 C ln / (7.8)

ı
We conclude that the lower bound on the number of the training examples that
are needed if the (
, ı)-requirements are to be satisfied grows here exponentially in
the number of attributes, n. Such growth is prohibitive for any realistic value of n.
For instance, even if we add only a single attribute, n C 1, the value of ln jHj will
double because ln jHj D 2nC1 ln 2 which is twice a much as 2n ln 2. And if we add
ten attributes, n C 10, then the value of ln jHj increases a thousand times because
ln jHj D 2nC10 ln 2 D 210 2n ln 2 D 1024 2n ln 2.
This observation is enough to convince us that a classifier in this general form is
not PAC-learnable.
A Word of Caution We must be careful not to jump to conclusions. As pointed
out earlier, the derivation of Inequality (7.5)—a worst-case analysis of sorts—
relied on quite a few simplifying assumptions that render the obtained bounds
very conservative. In reality, the number of the training examples needed for the
induction of a reliable classifier is much lower than that indicated by our “magic
formula.”
For the engineer seeking to choose the most appropriate learning technique, the
main contribution of Inequality (7.5) is that it helps him compare PAC-learnability
of classifiers constrained by different “languages.” For instance, we have seen that a
7.3 Some Practical and Theoretical Consequences 143
conjunction of attribute values can be learned from a reasonably sized training set,
whereas a general concept defined by any boolean function usually cannot.
Some other corollaries of this analysis will be the subject of Sect. 7.3.
• What is the size of a hypothesis space consisting of conjunctions of attribute
values? Substitute this size to Inequality (7.5) and discuss the result.
• Can you find the value of jHj for some other class of classifiers?
• Explain, in plain English, why a boolean function in its general form is not PAC-
learnable.
7.3 Some Practical and Theoretical Consequences
The theoretical analysis from the first section of this chapter offers a broader
perspective of the learning task as well as clues whose significance cannot be
overstated. Their benefits range from purely intellectual satisfaction enjoyed by a
theoretician to practical guidelines appreciated by the most down-to-earth engineer.
Let us take a quick look at some of them.
Bias and Learnability Section 7.2 investigated PAC-learnability of classes whose
descriptions are limited to conjunctions of attribute values. A constraint of this kind
is called a bias for a specific type of class.
Allowing for some attributes to be ignored as irrelevant, we calculated the size of
the hypothesis space thus defined as jHj D 3n . If we strengthen the bias by insisting
that every single attribute must be involved, we will reduce the size of the hypothesis
space: jHj D 2n . This is because every attribute in the class-describing expression
is either true or false, whereas in the previous case, a third possibility (“ignored”)
was permitted.
Many other biases exist, some limiting the number of terms in the conjunction,
others preferring a disjunction or some pre-specified combination of conjunctions
and disjunctions, or imposing yet another constraint. What matters, for our discus-
sion, is that each bias is likely to result in a different size of the hypothesis space.
And, as we have seen, this size affects learnability.
No Learning Without a Bias In the absence of any bias, allowing for any boolean
function, we have established that the value of ln jHj grows exponentially in the
number of attributes, ln jHj D 2n ln 2, which means that the class is in this most
general form not PAC-learnable. The explanation is simple. The unconstrained
hypothesis space is so vast that there is a reasonable chance that one may find
a classifier that labels correctly the entire training set, and yet exhibits poor
performance on future examples. Put bluntly, the induced classifier cannot be
trusted.
It is in this sense that we say, with a grain of salt, that “there is no learning without
a bias.”
The thing to remember is that the machine-learning adventure can only succeed
if the engineer constrains the hypothesis space by a meaningful bias. It stands to
reason, however, that this bias should not be misleading. The reader will recall that
our analysis assumed that the hypothesis space does contain the solution, and that
the examples are noise-free.4
Occam’s Razor Quite often, the engineer has an opportunity to choose from two
or more biases. For instance, the class to be learned is described by a conjunction
of attribute values, in which each single attribute plays a part. A weaker bias that
permits the absence of some attributes from the conjunctions also includes the case
where zero attributes are absent, and is therefore correct, too. In the former case, we
have ln jHj D n ln 2; and in the latter, ln jHj D n ln 3, which is bigger. We thus have
two correct biases, each defining a hypothesis space of a different size. Which of
them to prefer?
The attentive reader no doubt already knows the answer. In Inequality (7.5), the
number of training examples needed for successful learning depends on ln jHj. A
lower value of this term indicates that fewer examples are needed if we want to
satisfy the given .
; ı)-requirements. The engineer will thus benefit by choosing the
bias whose hypothesis space is smaller.
By the way, scientists have been using a similar rule for centuries: in a situation
where two different hypotheses can explain a certain phenomenon, it is assumed that
the simpler one has a higher chance of success. The fact that this principle, Occam’s
Razor, has been named after a scholastic theologian who died in the fourteenth
century indicates that the use of this rule pre-dates modern science. The word razor
is here to emphasize that, when formulating a hypothesis, we better slice away all
unnecessary information.
That mathematicians have now been able to find a formal proof of the validity of
this principle in the field of machine learning, and even to quantify the scope of its
utility, is a remarkable achievement.
Irrelevant and Redundant Attributes In the types of classes investigated in
Sect. 7.2, the lower bound on the number of examples, m, depended on the number
of attributes, n. For instance, in the case of conjunctions of attribute values, we now
know that ln jHj D n ln 3; and the number of examples needed to satisfy given
.
; ı/-requirements thus grows linearly in n.
4
Analysis of situations where these requirements are not satisfied would go beyond the scope of an
introductory textbook.
7.4 VC-Dimension and Learnability 145
The same result enables us to form an opinion about how a class learnability
might be affected by the presence of irrelevant or redundant attributes. The higher
the number of these attributes, the higher the value of n, which means that more
training examples have to be provided if we want to satisfy the .
; ı/-requirements.
The lesson is clear: whenever we have a chance to identify (and remove) the less-
then-useful attributes, we better do so.
These considerations also explain why it is so difficult to induce a good
classifier in domains where only a tiny percentage of attributes carry the relevant
information—as is the case in text classification.
• What is meant by the statement, “there is no learning without a bias”? Con-
versely, under what circumstances will the engineer’s bias tend to hurt the
learning algorithm’s chances of success?
• Explain the meaning of the term, Occam’s Razor. In what way did mathemati-
cians provide a solid ground for this—formerly philosophical—principle?
• What does Inequality (7.5) tell us about the role of irrelevant and redundant
attributes?
7.4 VC-Dimension and Learnability
So far, we have only dealt with domains where the number of hypotheses is finite.
Let us now turn our attention to domains where the hypothesis space is infinite,
which is often inevitable in applications where at least some of the attributes are
continuous-valued. In this situation, it would appear that comparing the learnability
of two classes of classifiers is all but impossible because both of them have infinite
jHj anyway.
To begin with, the reader will agree that some classes of classifiers (say,
polynomial) are more flexible than others (say, linear). Surely this flexibility is likely
to have some impact on learnability? Indeed it does. Let us take a quick look at how
theoreticians quantify learnability under this situation.
Shattered Training Set Consider the three examples in the two-dimensional space
depicted on the left of Fig. 7.1. No matter how we distribute the positive and negative
labels among them, we will always be able to find a linear classifier that separates
the two classes (the classifiers are shown in the same picture). We say that this
particular set of examples is “shattered” by the class of linear classifiers.
Fig. 7.1 The set of the three points on the left is “shattered” by a linear classifier. The set of the
three points on the right is not shattered by a linear classifier because no straight line can separate
the point in the middle from those on the sides
This, however, is not the case of the three points on the right. As we can see,
these points find themselves all on the same line. If we label the one in the middle
as positive and the other two as negative, no linear classifier will ever succeed in
separating the two classes from each other. In other words, this particular set of
examples is not shattered by the linear classifier.
A different class of classifiers will have a different power to shatter a given
set of examples. For example, a parabola (a special kind of a quadratic, and thus
polynomial function) will shatter the three aligned points on the right; it will even
shatter four points that do not lie on the same line. And other classes of classifiers,
such as high-order polynomials, will shatter any realistically sized set of examples.
Vapnik-Chervonenkis Dimension The Vapnik-Chervonenkis dimension (usually
abbreviated as VC-dimension) of a given class of classifiers is defined as the size of
the largest set of examples shattered by this class.
We have seen that, in a two-dimensional space, a linear classifier fails to shatter
three points that are all on the same line, but that the linear classifier does shatter
them if they do not lie on the same line. At the same time, four points, no matter how
we arrange them in a plane, can always be labeled in a way that makes it impossible
to separate positive examples from the negative linearly. Since the definition says,
“the largest set of examples shattered by this class,” we are bound to conclude that
the VC-dimension of a linear classifier in a two-dimensional space is VCL D 3.
The point to remember is that the value of the VC-dimension reflects—and
quantifies—the geometrical properties of the given class of classifiers.
Learnability in Continuous Domains The concept of VC-dimension makes it
possible for us to deal with learnability in continuous domains. Let us give here
a slightly modified version of a famous theorem, omitting certain technicalities that
are irrelevant for our specific needs:
If the VC-dimension, d, of a classifier class, H, is finite, then an error rate below
can be achieved with confidence 1 ı if the target class is identical with some
hypothesis h 2 H, and if the number of the training examples, m, satisfies the
following inequality:
7.4 VC-Dimension and Learnability 147
Table 7.2 VC-dimensions Hypothesis class VC-dimension

for some hypothesis classes
in Rn Hyperplane nC1
Hypersphere nC2
Quadratic .nC1/.nC2/
2
!
r-Order polynomial nCr
r
4 2 8d 13
m max. log ; log / (7.9)

ı

Note that this means that the lower bound on m is either ( 4

log 2ı ) or ( 8d

log 13

/,
whichever is greater.
The engineer interprets this result as telling him that he can then trust any
classifier which correctly classifies the entire training set of size m, regardless of
the machine-learning algorithm that has induced the classifier.
Note that the number of examples necessary for PAC learning grows linearly in
the VC-dimension. This, however, is no reason to rejoice. Thing is, VC-dimensions
of many realistic classes of classifiers have a way of growing very fast with the
growing number of attributes—see below.
Some Example VC-Dimensions Table 7.2 lists some VC-dimensions for linear
and polynomial classifiers. The reader may want to contemplate the inevitable
trade-off: more complex classes of classifiers are more likely to contain the correct
solution, but the number of training examples needed for success increases so
dramatically that a classifier from this class cannot be deemed learnable.
Indeed, in the case of higher-order polynomials, the demands on the training-set
size become all but prohibitive. For instance, the VC-dimension of a second-order
polynomial in 100 dimensions (a fairly normal number of attributes) is as follows:
102 101
dD D 5050
21
This is much larger than that of a linear classifier, but perhaps still acceptable. At
any rate, the value is sufficiently high to make the first term in Inequality (7.9) small
enough (by comparison) to be ignored.
If, however, we increase the polynomial’s order to r D 4, the VC-dimension will
become all but prohibitive:
104 103 102 101

dD D 4;598;126
4321
Polynomials of higher order therefore better be avoided. On the other hand,

the VC-dimensions of neural networks and decision trees (see Chaps. 5 and 6) are
known to be more affordable—which is why these classifiers are preferred.
A Word of Caution Just as in the case of Inequality (7.5), this one (Formula (7.9))
is the result of a worst-case analysis during which certain simplifying assumptions
were made. The results therefore should not be interpreted as telling us how many
examples are needed in any concrete application.
Importantly, in realistic applications, examples are not labeled arbitrarily. Since
the examples of the same class are somehow similar to each other, they tend to be
clustered together in the instance space.
• How does the number of examples that are necessary for the .
; ı/-requirements
grow with the increasing VC-dimension, d?
• What is the VC-dimension of a linear classifier in the two-dimensional continu-
ous space?
• What is the VC-dimension of an n-dimensional polynomial of the r-th order?
Why does the high value of this VC-dimension mean that polynomial classifiers
may not be a good choice?
• The Computational Learning Theory has been able to establish certain limits on
the number of the training examples needed for successful classifier induction.
These limits depend on two fundamental parameters: the threshold on error rate,

, and the upper bound, ı, on the probability that m training examples will
succeed in eliminating all classifiers whose error rate is above
.
• If all attributes are discrete, the minimum number of examples needed to satisfy
the .
; ı/-requirements is determined by the size, jHj, of the given hypothesis
space. Here is the classical formula:
1 1
m> .ln jHj C ln / (7.10)

ı
• In a domain where some attributes are continuous, the minimum number of
examples needed to satisfy the .
; ı/-requirements is determined by the so-called
VC-dimension of the given class of classifiers. Specifically, if the value of the
7.6 Exercises and Thought Experiments 149
VC-dimension is denoted by d, then the number of the needed examples is

determined by the following inequality:
4 2 8d 13
m max. log ; log / (7.11)

ı

• Inequalities (7.5) and (7.9) have been found by a worst-case analysis. In reality,
therefore, much smaller training sets are usually sufficient for the induction of
high-quality classifiers. The main benefit of these formulas is that they help us
compare learnability in different machine-learning paradigms.
Historical Remarks The principles underlying the idea of PAC learning were
proposed by Valiant [91]. The classical paper on what later came to be known as
VC-dimension is Vapnik and Chervonenkis [94]; somewhat more elaborate version
was later developed by Vapnik [92]. The idea to apply VC-dimension to learnability,
and to the investigation of Occam’s Razor is due to Blumer et al. [6]. Tighter bounds
were later found, for instance, by Shawe-Taylor et al. [86]. VC-dimensions of linear
and polynomial classifiers can be derived from the results published by Cover [18].
Readers interested in learning more about the Computational Learning Theory will
greatly benefit from the excellent (even if somewhat older) book by Kearns and
Vazirani [41].
7.6 Exercises and Thought Experiments
The exercises are to solidify the acquired knowledge. The ambition of the suggested
thought experiments is to let the reader see this chapter’s ideas in a different light
and, somewhat immodestly, to provoke his or her independent thinking.
Exercises
1. Suppose that the instance space is defined by the attributes used in the “pies”
domain from Chap. 1. Determine the size of the hypothesis space if the classifier
is to be a conjunction of attribute values. Consider both cases: the one that
assumes that some attributes might be ignored as irrelevant (or redundant), and
the one that insists that all attributes must take part in the conjunction.
2. Return to the case of conjunctions of boolean attributes from Sect. 7.2. How many
more examples will have to be used (in the worst-case analysis) if we change the
required error rate from
D 0:2 to
D 0:05? Conversely, how will the size of
the necessary training set be affected by changes in ı?
3. Again, consider the case where all attributes are boolean, and the classifier has the
form of a conjunction of attribute values. What is the size of the hypothesis space
if the conjunction is permitted to involve exactly three attributes? For instance,

here is one conjunction from this class:
att1 = true AND attr2 = false AND att3 = false
4. Consider a domain with n D 20 continuous-valued attributes. Calculate the VC-
dimension for a classifier that has the form of a quadratic function, and compare
it with that of a third-order polynomial.
Second, suppose that the engineer has realized that half of the attributes are
irrelevant. Having removed them, he now has n D 10. How will this reduction
affect the VC-dimensions of the two classifiers?
5. Compare the PAC-learnability of the boolean function involving 8 attributes
with the PAC-learnability of a quadratic classifier in a domain with 4 numeric
attributes
1. A certain role in Inequality (7.5) is played by ı, a term that quantifies the

probability that a successful classifier will be induced from the given training
set. Under what conditions can the impact of ı be neglected?
2. From the perspective of PAC-learnability, is there a difference between irrelevant
and redundant attributes?
3. We have seen that a classifier is not PAC-learnable in the absence of a bias.
The bias, however, many not be known. Suggest a learning procedure that would
induce a classifier in the form of a boolean expression in this case. (Hint: consider
two or more alternative biases.)
4. In the past, some machine-learning scientists considered the idea of converting
continuous attributes into discrete ones by a process called discretization. By
this they meant dividing the range of attribute values into intervals, each interval
treated as a boolean attribute (the given numeric value either is or is not in the
given interval).
Suppose you are considering two ways of dividing the range Œ0; 100. The
first consists of two subintervals, Œ0; 50, Œ51; 100, and the second consists of
ten equally sized subintervals: Œ0; 10; : : : Œ91; 100. Discuss the advantages and
disadvantages of the two options from the perspective of PAC-learnability.
Chapter 8
A Few Instructive Applications
For someone wishing to become an expert on machine learning, mastering a handful

of baseline techniques is not enough. Far from it. The world lurking behind a
textbook’s toy domains has a way of complicating things, frustrating the engineer
with unexpected obstacles, and challenging everybody’s notion of what exactly the
induced classifier is supposed to do and why. Just as in any other field of technology,
success is hard to achieve without a healthy dose of creativity.
Some practical experience helps, too, either your own, or at least of those who
have tried and succeeded (or failed) before you. And this is what this chapter wants
to offer. Using a few carefully selected case studies, it will acquaint you with some
issues typically encountered in realistic applications—and with practical ways of
dealing with them.
8.1 Character Recognition
The techniques from the previous chapters target recognition skills that are too
difficult to be hard-coded in a computer program, but can be conveyed by means
of pre-classified training examples. The ability to read text, even hand-written
text, belongs to this category. Applications are legion: automated pre-processing
of various forms, software for converting a text scribbled on a notepad into a format
to be used by a text editor, and various newspaper-digitization programs. A mere
generation or two ago, an undertaking of this kind was judged so ambitious as to be
almost unrealistic; today, no one finds it extraordinary.
The Task To describe each character in a way that facilitates its recognition by
a computer is not easy. To get the idea, just try to explain, in plain English, how
to distinguish digit “5” from digit “6,” or what constitutes the differencebetween a

DOI 10.1007/978-3-319-63913-0_8
152 8 A Few Instructive Applications
Fig. 8.1 A simple way to

convert the image of a
hand-written character into a
vector of 64 continuous a 6−by−4 matrix of
attributes, each giving the numeric attributes,
mean intensity of the each giving mean
corresponding field
intensity of a field
hand-written “t” and “l.” You will be surprised how difficult this is. And to convert
the plain-English explanation to a code that a computer can understand is harder
still.
This is why machine learning has been asked to help. If we prepare a collection of
training examples, and describe them by pre-classified attribute vectors, then surely
some of the techniques described in the previous chapters can induce the requisite
classifier. The process is so simple as to seem almost trivial. Let us take a look at
some of its crucial aspects.
Examples and Attribute Vectors To begin with, the engineer has to define the
attributes to describe the examples. Figure 8.1 illustrates one possibility. The first
step identifies a rectangular region in which the digit is located; the second divides
this region into 6 4 D 64 equally sized fields, each characterized by a continuous
attribute whose value gives the field’s mean intensity. More ink means the field
reflects less light, and thus results in a lower value of the attribute, and, conversely,
the value is maximized if the field contains no ink at all. Of course, this is just the
principle. There is no reason why there should be only 64 attributes. Indeed, the
higher the level of detail the program is expected to discern, the smaller the size of
the fields it needs to rely on, and thus the greater the number of attributes that will
be used to describe the examples.
In a realistic application, the classifier will have to recognize not just one isolated
character (such as the “3” in Fig. 8.1), but rather to “read” a whole text consisting of
many such characters. This generalization, however, does not represent any major
complication. As long as the individual characters can be isolated, and then treated
separately, the same principle of describing each of them by an attribute vector can
be applied.
Nowadays it is possible to download datasets of hundreds of thousands of hand-
written characters, some described in the way presented above, some relying on
other approaches.
Choosing the Classifier Now that we know how the training examples are going
to be described, we are ready to proceed to the choice of the induction technique.
8.1 Character Recognition 153
The first thing to be considered is that, in the attribute-vector obtained by the

mechanism from Fig. 8.1, only a small percentage of the attributes (if any) are likely
to be irrelevant or redundant, which means that this is not an issue to worry about
(unless the number of attributes is increased way beyond those shown in Fig. 8.1).
Another aspect guiding our choice of an appropriate machine-learning paradigm
is the fact that we do not know whether the classes are linearly separable, and
therefore hesitate to use linear classifiers. Finally, the classifiers need not be capable
of offering explanations. If the intention is to read and convert to a text editor a long
text, the user will hardly care to know the exact reason why a concrete character was
classified as a “P” and not as a “D.”
Based on these observations, the simple and easy-to-implement k-NN classifier
looks like a good candidate. A cautious engineer will perhaps be concerned about
the computational costs incurred in a domain with hundreds of thousands of training
and testing examples. But as long as the number of attributes is moderate, these
costs are unlikely to be prohibitive. From the practical point of view, a reasonable
limit on what constitutes “prohibitive costs” will be determined by the computations
associated with the isolation of the individual characters and the conversion of
their images to attribute vectors. As long as these are comparable with the costs
of classification (and they usually are), the classifier is affordable.
And indeed, the nearest-neighbor classifier is the most common choice, in this
application, typically exhibiting an error rate of less than 2%. In some really illegible
hand-writings, the error rate will of course be higher. But then, we should not be
too harsh on the innocent machine, knowing as we do that even an experienced
pharmacist finds it difficult to read certain hand-written prescriptions.
The Number of Classes In a domain where all characters are capitalized, the
induced classifier is to discern 10 digits and 26 letters, which amounts to 36
classes. If both lowercase and uppercase letters are allowed, the total increases to
10 C 2 26 D 62 classes, and to these we may have to add special characters such
as “?,” “!,” “$,” and so on. This relatively high number of classes is not without
consequences, and these deserve our attention.
The most immediate concern is the induced product’s evaluation. Mere infor-
mation about the error rate is somewhat inadequate here. Thus the performance of
a classifier that correctly identifies 98% characters may appear good enough, even
impressive; what this value fails to tell us, though, is how the errors are distributed.
Typically, some characters will be correctly identified most of the time, while others
pose difficulties—and as such, deserve further attention. For instance, certain pairs
of similar characters tend to be mutually confused; the practically minded engineer
then wants to know which pairs so as to mitigate the problem by providing additional
training examples for the “difficult” classes.
Moreover, some letters will be less common than others. In a situation of this
kind, it is known that the rarer classes get “overlooked” by the classifier-inducing
algorithms unless special precautions have been taken. Section 10.2 will have more
to say about this issue.
The Classifier Should Be Allowed to Reject an Example To decipher a person’s

handwriting is far from easy. Certain letters are so ambiguous as to make the reader
shrug his shoulders in despair. Yet the k-NN classifier from Chap. 3 is undeterred:
it always finds a nearest neighbor, and then simply returns its class, no matter how
arbitrary this class is.
Practical experience shows this circumstance to be harmful because the costs of
getting the wrong class can be greater than the costs of not knowing the class at all.
Thus in an automated reader of postal codes, an incorrectly read digit can result in
the letter being sent to a wrong destination, which, in turn, may cause great delay
in delivery. On the other hand, if the classifier does not give any answer, a human
employee will have to do the reading. The costs of manual processing may then be
lower than the costs of getting the wrong address.
We have convinced ourselves that the classifier should be implemented in a way
that makes it possible to refuse to classify an example if the evidence supporting
the winning class is insufficient. The simplest way of doing so in the context of
the k-NN classifier is to require a certain minimum margin between the number
of votes supporting the winner and the number of votes supporting the runner-up.
For instance, if the winning class in a 7-NN classifier receives only four votes as
compared to the three votes supporting another class, the example is rejected as
ambiguous.
Something similar is easy to accomplish also in some other paradigms such
as the Bayesian classifiers or neural networks: the classifier simply compares the
probabilities (or output signals) of the two most likely classes, and rejects this
example if the difference does not exceed a predefined threshold.
Error Rate Versus Rejection Rate A classifier that rejects ambiguous examples
will surely reduce its error rate; on the other hand, excessive reluctance to classify
will not be beneficial, either. What if all examples are rejected? The error rate then
drops to zero—and yet the user will question the tool’s practical utility.
The lesson is, the engineer needs to consider the trade-off between the rejection
rate and the error rate. True enough, increasing the former is likely to reduce the
latter; but overdoing it may render the classifier useless.
• Explain how to describe hand-written characters by attribute vectors.
• What aspects did we consider (and why) when looking for the most appropriate
machine-learning tool to be used here?
• What is the immediate consequence of the relatively high number of classes in
this domain? Why is here the error rate unable to give the full picture?
8.2 Oil-Spill Recognition 155
• Why should the classifier be allowed to refuse to classify certain examples?

Discuss the trade-off between error rate and rejection rate; comment on the
interplay between performance and utility.
8.2 Oil-Spill Recognition
Figure 8.2 shows a radar image of the sea surface as taken by a satellite-born device.
Against the grayish background, the reader can see several dark regions of the most
varied characteristics: small or large, sharp or barely discernible, and of all possible
shapes. What interests us here primarily is the sharp and elongated object in the
vicinity of the upper-right corner: an oil spill, illegally dumped by a tanker’s captain
who has chosen to get rid of the residues in its bilges in the open sea rather than
doing so in a specially designed terminal as required by the law. As such, this
particular oil spoil is of great interest to the Coast Guard.
For all relevant sea-surface areas (close to major ports, for example), satellites
take many of such “snapshots” and send them down to collaborating ground stations.
Here, experts pore over these images in search for signs of illegal oil spills.
Whenever they detect one, an airplane is dispatched and verifies the suspicion by
a spectrometer (which is unavailable to the satellite), collects evidence, and perhaps
even identifies the perpetrator.
Unfortunately, human experts are expensive—and not always available. They
may be on holidays, on a sick leave, or not present for any number of other reasons.
Besides, in view of the high number of images, the work is tedious, and prone to
human errors. This is why the idea came up to develop a computer program to
automate the oil-spill recognition process.
Fig. 8.2 A radar image of a sea surface. The “wiggly” elongated dark region in the upper-right
corner represents environmental hazard: an oil spill
The picture shown in Fig. 8.2 has been selected out of many, the main criterion
being its rare clarity. Indeed, the oil spill it contains is so different from the other
dark regions that even an untrained eye will easily recognize it as such. Even so,
the reader will find it difficult to specify the oil-spill’s distinguishing features in a
manner that can be used in a computer program. In the case of more realistic objects,
the task will be even more difficult. At any rate, to hard-code the oil-spill recognition
ability is quite a challenge.
Again, machine learning got its chance, the intention being to let the machine
develop the requisite skills automatically, by induction from training examples. The
general scenario of the adventure can be summarized into the following steps.
1. collect a set of radar images containing oil spills;
2. use some image-processing software capable of identifying, in these images, the
dark regions of interest;
3. ask an expert to label the oil spills as positive examples, and the other dark
regions (so-called “look-alikes”) as negative examples;
4. describe all examples by attribute vectors, and let a machine-learning program
induce the requisite classifier from the training set thus obtained.
As in the previous application, we will try to glean some useful lessons by taking
a closer look at certain critical aspects of this undertaking.
Attributes and Class Labels State-of-the-art image-processing techniques easily
discover dark regions in an image. For these to be used by the machine-learning tool,
we need to describe them by attributes that are hoped to capture those aspects that
distinguish spills from “look-alikes.” Preferably, their values should be unaffected
by the given object’s size and orientation.
The attributes that have been used in this project include the region’s mean
intensity, average edge-gradient (which quantifies the sharpness of the edges), the
ratio between the lengths of the object’s minor-axis and major-axis, variance of the
background intensity, variance of the edge gradient, and so on. All in all, more than
forty such attributes were selected in a rather ad hoc manner. Which of them would
really be useful was hard to tell because experts were unable to reach consensus
about the attributes’ respective relevance and redundancy. The final choice was left
to the machine-learning software.
Labeling the training examples with classes was not any easier. The objects in
Fig. 8.2 were easy to categorize; in other images, they were much more ambiguous.
On many occasions, the best the expert could say was, “yes, this looks like an oil
spill” or, “I rather doubt this is what we are looking for.” The correctness of the
selected class labels was thus uncertain, resulting in class-label noise. The presence
of class-label noise reduces our expectations: the classifier can be only as good as
the data that have been used during its induction.
Choosing the Classifier The examples were described by some forty attributes.
Among these, some, perhaps most, were suspected of being either irrelevant or
redundant. This is an important circumstance; the reader will recall that the presence
of irrelevant and redundant attributes makes some classifiers underperform. In this
8.2 Oil-Spill Recognition 157
particular project, the problem was side-stepped by first inducing a decision tree,
and then eliminating all attributes that never appeared in any of the tree’s tests. This
was quite logical. After all, the choice of which attributes to include in the tree has
been made based on the attributes’ information contents. Since information contents
of irrelevant attributes is low, this method of identifying them is quite reliable. Also
redundant attribute can thus be eliminated, at least to some extent.
When the k-NN classifier was applied to examples described by attributes that
“survived” this decision-tree-based elimination process, the classification perfor-
mance turned out to be acceptable. The oil-spill problem was thus solved with a very
simple machine-learning technique—which is good: having a choice, the engineer
is always well advised to give preference to the simpler tool.
The decision to use the k-NN classifier was also driven by another important
consideration. Since this is typical of many realistic applications, let us discuss it in
some detail.
Cost of Errors Performance evaluation is here not as straightforward as it was in
the previous chapters. For one thing, error rate can be a fairly misleading indicator
in domains where each type of error carries a different penalty.
This is manifestly the case in the oil-spill adventure. Here, a false positive results
in an aircraft being unnecessarily dispatched to a suspicious though “innocent”
region, and this means a waste of time and resources (note that the costs are not
only financial, but also moral because these kind of failures tend to undermine the
user’s trust in the classifier). On the other hand, a false negative means an undetected
environmental hazard whose consequences (financial, environmental, and political)
are hard to predict. In view of all this, the reader will agree that the two types of cost
are of such a different nature that it is almost impossible to compare them. This, of
course, makes it difficult to specify the project’s goal.
Experiments on pre-classified testing data indicated that most of the errors made
by the induced classifier were of the false-positive kind (i.e., false alarms). As for
false negatives, these were relatively rare. But then: was this good, or should this
observation be taken as a signal of a need to modify the classifier in order to change
the ratio of the two types of errors?
The question cannot be answered in isolation from the application’s momentary
needs. Financial constraints may force the user occasionally to accept the risk
of environmental hazard, simply because the budget can no longer tolerate false
alarms. The user then wants to reduce the frequency of false positives even if
this means to pay the price of an increased number of undetected oil spills (false
negatives).
However, the situation will change in more prosperous times when the user who
does not want to miss an oil spill is prepared to accept higher frequency of false
positives for the sake of making false negatives rare.
Leaning Toward One or the Other Class In view of these trade-offs, it is
necessary to give the user the opportunity to adjust the classifier’s behavior so as
to modify the frequency of one or the other type of error.
As already mentioned, this project relied on the k-NN classifier where this
requirement is easy to satisfy: the trick consists in manipulating the margin between
the number of votes supporting either of the two classes. For the sake of illustration,
suppose the 7-NN classifier is used. Here, the number of false positives can be
reduced if we instruct the classifier to label as positive only those examples where,
say, at least five of the nearest neighbors are positive; any example that fails to satisfy
this condition is deemed negative. Conversely, if we desire to lower the frequency
of false negatives, we tell the classifier to return the positive label whenever, say,
at least three of the nearest neighbors are positive. The user’s preference for either
type of error is thus expressed in terms of the number of votes that are necessary for
the example to be labeled as positive.
• How did the engineers deal with the fact that many attributes were redundant or
irrelevant? What did they identify these less-then-useful attributes?
• What can be said about the reliability of the class labels in the training set? What
does it mean for the classifier’s expected performance?
• Discuss the respective costs, in this domain, of the two types of error: false
positive versus false negative. Can they be compared using the same units? Why
does the user need a mechanism to reduce one type of error at the cost of the
other?
• Explain the essence of the mechanism that enables the k-NN classifier to increase
or reduce either of the two errors.
8.3 Sleep Classification
Throughout the night, we go through different sleep stages such as deep, shallow,
or rapid-eye movements (REM, this is when we dream). To identify these stages in
a concrete sleeping subject, advanced instrumentation is used: an electrooculogram
to record eye movements, an electromyogram to record muscle contractions, and
contact electrodes attached to the scalp to record the brain’s neural signals. Based
on the readings of all these instruments, a medical expert can decide what stage the
sleeping subject is in at any given moment, and can even draw a hypnogram such as
the one shown in Fig. 8.3.
Note that the deepest sleep stage occurs here only three times during the 8-h
sleep, and that it usually does not last long. Note also the move stage; this occurs,
for instance, when the subject turns from one side to another, moves the head or an
arm, and the like.
8.3 Sleep Classification 159
wake
move
REM
stage 1
stage 2
stage 3/4
hours of sleep
2 4 6 8
Fig. 8.3 An example hypnogram that records the sleep stages experienced by a subject during an
8-h sleep
Why Is It Important Medical practice needs to be able to recognize specific sleep

stages. A case in point is the so-called sudden infant death syndrome (SIDS): an
infant dies without any apparent cause. A newborn suspected of being in danger
has to be watched, in a hospital, so that resuscitation can be started immediately.
Fortunately, SIDS is known almost always to occur during the REM stage. This
means that it is not necessary to watch the patient all the time, but only during this
period of increased risk. For instance, a device capable of recognizing the onset of
the REM stage might alert the nurse that more attention might be needed during the
next five of so minutes.
The hypnogram, in turn, is a useful diagnostic tool because the distribution of the
sleep stages during the night may indicate specific neural disorders such as epilepsy.
Why Machine Learning To draw the hypnogram manually is a slow and tedious
undertaking, easily taking 3–5 h of a highly qualified expert’s time. Moreover, the
expert is not always available. This is why efforts have been made to develop
a computer program capable of identifying the individual sleep stages based on
observed data, and, hopefully, even to draw the hypnogram.
To be able to help, the computer needs a description of the individual stages. Such
description, however, is difficult to obtain. Medical experts rely on skills obtained
in the course of long training, and they use features and indications that are too
subjective to be converted to a computer program.
This motivated an attempt to induce the classifier from pre-classified data.
Specifically, the data-acquisition process divided the 8-h sleep period into 30-s
samples, each of them treated as a separate training example. All in all, a few hours’
sleep thus provided hundreds of training examples. Diverse instruments than provide
data to describe each 30-s sample.
Attributes and Classes Again, the first task was to remove attributes suspected of
being irrelevant or redundant. In the previous application, oil-spill recognition, this
removal was motivated by the intention to increase the performance of the k-NN
classifier (which is known to be sensitive to their presence). In sleep classification,
another reason comes to the fore: the physician wants to minimize the number of
measurement devices attached to the sleeping subject. Not only does their presence
make the subject feel uncomfortable; they also disturb the sleep, and thus interfere
with the results of the sleep analysis.
As for the class labels, these are even less reliable than in the oil-spills domain.
The differences between “neighboring” (similar) sleep stages are so poorly defined
that any two experts rarely agree on more than 70–80% of the class labels. No
wonder that the training set contains a lot of class-label noise, and the low quality
of the data imposes a limit on the minimum error rate that any realistic classifier
induced from the data can achieve.
The Classifier and Its Performance The classifier employed in this particular
case combined decision trees with a neural network in a manner whose details are
unimportant of our needs here. Suffice it to say that the classifier’s accuracy on
independent data indeed achieved those 70–80% observed in human experts, which
means that the natural performance limits have been reached. It is perhaps worth
noting that plain decision trees were a few percent weaker than that.
This said, it is important to understand that classification accuracy does not give
the full picture of the classifier’s behavior (similarly as in the OCR domain from
Sect. 8.1). For one thing, the classifier correctly recognized some of the seven classes
most of the time, while failing on others. Particularly disappointing was its treatment
of the REM state. Here, classification accuracy was in the range 90–95%, which,
at first sight, looked good enough. However, closer inspection of the training data
revealed that only less than 10% of all examples belonged to the REM class; this
means that a comparable classification accuracy could be achieved by a classifier
that says, “there is not a single REM example”—and yet this is hardly what the
engineers hoped to achieve.
We realize that this way of measuring performance is not without its limitations,
and that other criteria have to be found. These indeed exist. They will be discussed
in Chap. 11.
Improving Classification Performance by Post-processing The accuracy of the
hypnogram can be improved by post-processing whose nature relies on the domain’s
logic. Indeed, several rules of thumb can be used here. For instance, the deepest
sleep (stage 3/4) is unlikely to occur right after the REM stage, and stage 2
does not happen after move. Also, the hypnogram can be “smoothed out” by the
removal of any stage lasting only one 30-s period. Applying such rules in the course
of post-processing makes it possible to improve the hypnogram’s informational
value.
The lesson is worth remembering. In domains where the examples are ordered
in time, the classes of the individual examples may not be independent of those
preceding or following them. In this event, post-processing can improve the
classifier’s performance.
Proper Use of the Induced Classifier The original idea was to induce the classifier
from examples obtained from a few subjects, and then to classify future data using
this induced classifier instead of the much more expensive “manual” classification.
8.4 Brain–Computer Interface 161
This ambition, however, turned out to be unrealistic. Practical experience showed

that no “universal classifier” of sleep data could be obtained in this manner: a
classifier induced from one person’s data could not be used to draw a hypnogram
for another person without serious degradation in classification performance.1
This does not mean that machine learning is in this domain totally disqualified.
Far from it. However, the user needs to modify his or her expectations: instead
of being universal, the classifier will be induced separately for each subject. The
expert’s efforts are still significantly reduced. The following three-step scenario is
indicated:
1. The expert determines the class labels of a subset of the available examples, thus
creating a training set.
2. From this training set, a classifier is induced.
3. The induced classifier is used to classify the remaining examples, thus saving the
expert’s time.
• Why was it important to minimize the number of attributes in this domain? How
were the relevant attributes identified?
• The sleep-classification domain has seven classes. What does it mean for an
engineer trying to evaluate the induced classifier’s performance?
• In the hypnogram, the examples are ordered in time. How can this circumstance
be exploited in data post-processing?
• In what manner can machine learning reduce the burden imposed on someone
who seeks to classify available data?
8.4 Brain–Computer Interface
The muscle-controlling commands are issued at specific locations of motor cortex,

a relatively well-understood region of the cortex. Even the brain of many totally
paralyzed patients is fully capable of generating these commands; unfortunately, the
information fails to reach the muscles. The fact that these signals can be detected
by contact electrodes inspired a fantastic idea: can these signals, properly recorded
and interpreted, be used to control a cursor on a computer screen? If yes, then there
1
One can speculate that a different set of attributes might perhaps make this possible; the case
study reported here did not attempt to do so.
Fig. 8.4 Organization of the RS

WS CUE movement
experiment. After a “ready”
signal (WS) comes the CUE
(“left” vs. “right”). Once RS
appears on the screen, the
subject waits 1 s, then presses
the button indicated by the
CUE
0 1 2 3 4 5 6 7 8
Fig. 8.5 Arrangement of Cz
electrodes on the subject’s C3 C4
scalp. Only some of the
electrodes are actually used
(highlighted)
might be a way for the patient to communicate with the outside world, in spite of
being unable to speak, in spite of being unable even to move his or her pupils.
The exact nature of the motor commands is too complicated to be expressed by
a mathematical formula. But we can record the signals, describe them by attributes,
and label with the concrete motor commands. If we do so, we have a training set
from which a machine-learning program can induce the classifier.
The Training Examples In the specific case to be described in this section, only
two classes were considered: left and right. The training and testing examples
were obtained in the course of the following experiment. A subject was sitting in
front of a computer monitor. On the desk in front of him was a wooden board with
two buttons, one to be pressed by the left index finger, the other by the right index
finger, according to instructions displayed on the monitor (Fig. 8.4).
The whole scenario followed the time-line shown in the upper part of Fig. 8.5.
The contact electrodes attached to the scalp (see the bottom part of Fig. 8.5) record
the intensity of the neural signals during a certain period of time before the button
is pressed. The signals from each electrode are characterized by five numeric
attributes, each representing the power of the signal over a specific time interval
(a fraction of a second).
As indicated in the picture, only some (typically 11) of the electrodes were
actually used. For 11 electrodes, each represented by 5 attributes, this amounted
to 55 attributes. All in all, the subject “produced” a few hundred examples with
both classes equally represented. Since it was always known which button was
8.4 Brain–Computer Interface 163
actually pressed, the training set was free of class-label noise. As for the attributes,
the experimenters suspected many of them to be irrelevant.
Classifier Induction and Its Limitations Several machine learning paradigms
were experimented with, including multilayer perceptrons and nearest-neighbor
classifiers with attribute-value ranges normalized so as to avoid scaling-related
problems. Relevant attributes were selected by a decision tree, similarly as in the
applications discussed in the previous sections.
Typical error rates of these induced classifiers on testing data were in the range
20–30%, depending on the concrete subject. It is quite possible that better accuracy
could not be achieved: the information provided by the electrodes might have been
insufficient for this kind of experiment.
Importantly, just as in the sleep-classification domain, the classifier could only
be applied to the subject from whose training data it has been induced. All attempts
to induce a “general classifier” (to be used on any future subject) failed—the error
rates were so high as to make the classifications process look random.
Testing the Classifier’s Utility As in the previous domain, the error rate measured
on independent data does not give the whole picture; the classifier’s practical utility
depends on how it is actually employed. In our case, the real question is: will the
classifier succeed in sending the cursor to the correct location?
The question was answered by an experiment illustrated in Fig. 8.6. The subject
is still sitting in front of a computer monitor, with electrodes attached to the scalp.
However, the board with the two buttons has been removed. On the monitor, we can
see two rectangles, one on the left and one on the right. In the middle is a cursor in
the shape of a little cross.
When instructed to move the cursor, say, to the right, the subject only imagined he
was pushing the right button. The electrodes recorded the neural signals and passed
them to the classifier. Based on these, the classifier selected the side (the class, left
or right) to which the cursor is to be moved. The movement was very slow so as
to give the subject the opportunity to “correct” the wrong direction. To make this
possible, the neural signals were sampled repeatedly, each sample becoming one
example to be passed on to the classifier. Suppose the subject’s intention is to move
the cursor to the right. It is possible that the first sample will be misclassified as
Fig. 8.6 The classifier’s task

is to move the cross into
either the left box or the right
box, according to the
subject’s “thoughts”
left, sending the cursor in the wrong direction; but if the following samples are
correctly classified as right, the cursor will, after all, land in the correct rectangle.
What Error Rate Is Acceptable? Under the scenario just described, an error rate
of 30% (which seems quite high) turned out to be fully acceptable. Even when the
cursor occasionally did start in the wrong direction (on account of one “example”
being misclassified), there was still the chance of correcting the mistake before the
cursor reached its destination. As a result, the cursor almost always landed in the
correct box, even if in a somewhat hesitating manner in which it moved back and
forth before finding the right direction.
An impartial observer, uninformed about how the classifier was actually used,
rarely noticed anything was going wrong. The occasional errors on the part of the
classifier (manifested by the cursor’s back-and-forth movements) were sometimes
perceived as a sign of indecisiveness on the part of the subject, and not of the
classifier’s imperfection.
Only when the classifier’s error rate dropped well below those 30% did the cursor
miss its target frequently enough to cause disappointment. One can speculate that a
better measure of the classifier’s performance would in this case be the average time
needed for moving the cursor.
An Honest Historical Remark The classifier’s performance differed from one
person to another. On quite a high percentage of the subjects, the induced classifiers
exhibited an almost prohibitively high error rate. Also the fact that the classifier had
to be trained on the same patient on which it was supposed to be used was perceived
as a shortcoming. Further on, two classes are not enough. At the very least, also up
and down classes would be needed. On these, the experimental results were less
impressive. In the end, other methods of communication with paralyzed patients
came to be studied. Though successful as a machine-learning exercise, the project
did not satisfy the expectations of the medical domain.
• Discuss the experience made in this application: a classifier induced from the
data of one subject cannot be used to classify examples in another subject.
• Explain why the error rate of 30% was in this domain still deemed acceptable.
What other methods of measuring performance, in this application would you
recommend?
8.5 Medical Diagnosis 165
8.5 Medical Diagnosis
A physician attempting to find the cause of her patient’s complaints behaves a

bit like a classifier: based on certain attributes (the patient’s symptoms, results of
laboratory tests), she suggests a diagnosis—and a treatment. No wonder that, in the
early days of machine learning, many believed medical diagnosis to be its natural,
almost straightforward application.
Some Results The optimism was motivated by early studies such as the one whose
results are summarized in Table 8.1. The numbers give the classification accuracy
achieved by Bayesian classifiers, decision trees, and physicians on the same testing
set—patients described on attribute vectors. The four domains differ in the number
of classes and in the difficulty of the task (e.g., noise in the data, relevance and
reliability of the available information).
The values indeed seem to confirm machine-learning’s ability to induce classi-
fiers capable of competing with human experts; indeed, they seem to outperform
them. In the multi-class primary tumor domain, the margin is perhaps unim-
pressive, but in all the other domains, the machine learning classifiers appear to be
clearly better. It is only a pity that the table does not tell us more about the variation
in the results. At least standard deviations would help, here, but there are other ways
how to ascertain whether the results are statistically reliable—Chaps. 11 and 12 will
have more to say about the methods to be used here.
Are the Results Encouraging? The results seem impressive, but we must not jump
to conclusions. The first question to ask is whether the comparison was fair. Since
the examples were described by attributes, the data available to the machine-learning
tool, and to the induced classifier, could hardly be the same as used by the physician
who could rely also on subjective information that might not have been available to
the machine. In this respect, the human enjoyed a certain advantage, and this makes
the machine learning’s results all the more impressive.
On the other hand, the authors of the study admit that the physicians participating
in the study were not necessarily top specialists in the given field, and machine
learning thus outperformed only general practitioners. Another aspect worth our
attention is that this study was conducted in the 1980s when the diagnostic tools
were less sophisticated than those in use today. It is hard to tell who will benefit
more from modern laboratory tests: human experts or machine learning?
Table 8.1 Classification Bayesian Decision General

accuracy of two classifiers classifier tree practitioner
compared to the performance
Primary tumor 0.49 0.44 0.42
of physicians in four medical
domains Breast cancer 0.78 0.77 0.64
Thyroid 0.70 0.73 0.64
Rheumatology 0.67 0.61 0.56
Apart from Classification, Also Explanation Is Needed For medical diagnosis,

classification accuracy is not enough; more accurately, it is not enough for the
diagnosis to be correct. A patient will hardly agree with a surgery if the only
argument supporting this recommendation is, “this is what the machine says.” The
statement that “the machine is on average a few percentage points better than a
physician” is unlikely to convince the patient, either.
What a reasonable person wants is a convincing explanation of why the surgery is
to be preferred over a more conservative treatment. In the case of the domains from
Table 8.1, the doctor is usually able to itemize what evidence supports the concrete
diagnosis and why, and what treatment options are recommended in this particular
case and why. Unfortunately, the baseline machine learning techniques are not very
good at explaining their decisions (perhaps with the exception of decision trees).
Software for automated reasoning would be needed, here. This, however, is beyond
the scope of this book.
The Need for Measuring Confidence There is another problem to be aware
of. Most of the classifiers induced by the techniques treated in the previous
chapters only give the final verdict; for instance, “the example belongs to class 77.”
The physician (as well as the patient) needs to know more. For instance, is the
recommended class absolutely certain or is it only just a little more likely than some
other class? And if so, which other classes should be considered?
In other words, medical domains usually call for classifiers capable of telling us
how confident they are in the returned class. For example, such classifier should say
that “the example belongs to C1 with probability 0.8, and to C5 with probability 0.7.”
There probabilities are explicitly calculated by Bayesian classifiers, so these may be
a good choice; similar information can be obtained also from neural networks. Both
of these approaches, however, are rather unable to offer explanations.
On the other hand, decision trees are capable of offering some kind of explana-
tion. As for the confidence, this is only possible with additional functions that we
have not treated here.
Cultural Barriers There is one last reason why the encouraging results of the early
applications of machine learning to medical diagnosis failed to inspire followers.
With only minor injustice, they can be called cultural barriers. In a way, they are
understandable. Medical doctors surely did not like to be told how easily they would
be replaced by computers. No wonder that, in the early days of machine learning,
many of them were not eager to collaborate on the development of the requisite
software.
This, of course, is a misunderstanding. Machine learning is not meant to replace
a human expert. At its very best, is only offer advice; the final diagnostic decision
will always be the responsibility (even in legal terms) of the physician treating the
patient. Still, the value of the advice should not be underestimated. For instance, the
classifier can alert the doctor that additional, previously unsuspected, disorders my
accompany the current diagnosis (the patient often suffers from more than just one
problem at a time), and it can even point out the need to take specific additional
laboratory tests.
8.6 Text Classification 167
• As for the results mentioned in this section, why was it impossible to conclude
from them that the induced classifiers outperformed the human expert?
• Apart from classification accuracy, what is needed in medical diagnosis?
• Discuss the limitations of machine-based diagnosis. Suggest a reasonable way of
realistic application of machine learning in medical domains.
8.6 Text Classification
Suppose you have a vast collection of text documents, and you need to decide which
of them is relevant to a specific topic of interest. If there are really a great many
of them, there is no way you can do so manually. However, taking inspiration for
some of the above-described applications, you choose a manageable subset of these
documents, read them, assign the class labels to them (positive versus negative), and
induce a classifier that will then be used to classify the rest.
Attributes A common way to describe a document is by the frequency of the
words it contains. Each attribute represents one word, and its value represents the
frequency with which it appears in the text. For instance, in a document that is 983
words long, the term “classifier” may appear five times, in which case its frequency
is 5=983 D 0:005.
Since the vocabulary typically contains tens of thousands of words, the attribute
vectors will be very long. And of course, since only a small subset of the vocabulary
finds its way into the document, most attribute values will be zero. This being
somewhat unwieldy, simple applications prefer to work with only a subset of the
whole vocabulary, say, 1000 most common words.
Class Labels The class labels for the training examples may be difficult to
determine. Even if we want only something very easy such as to decide, for each
document, if it deals with computer science, whether an example clearly
belongs into this class may not be easy to decide because some documents are more
relevant to this category than others. For instance, a very clear-cut case will be a
scientific paper dealing with algorithm analysis; on the other hand, a less relevant
document will only mention in passing that some algorithms are more expensive
than others; and yet another document will be an article from a popular magazine.
One possibility to deal with this circumstance is to “grade” the class labels. For
instance, if the document is most certainly relevant, the class is 2; if it is only
somewhat relevant, the class is 1; and only if it is totally irrelevant, the class is 0.
The user can then decide what level of relevance is needed for the given application.
Typical Observations The great number of attributes makes the induction compu-
tationally expensive. Thus in the case of decision trees, where one has to calculate
the information gain separately for each attribute, this means that tens of thousands
of attributes need to be investigated for each of the tests, and these calculations
can take a lot of time, especially if the number of examples is high. Upon some
thought, the reader will agree that Bayesian classifiers and neural networks are quite
expensive, too.
One can suggest that induction will be cheaper in the case of the k-NN classifier
or a linear classifier. Here, however, another difficulty comes to the fore: for each
given topic, the vast majority of the attributes will be irrelevant, a circumstance that
reduces the utility of both of these paradigms.
The Problem of Multi-Label Examples The applications discussed in the pre-
vious sections assumed that each example is labeled with one and only one class.
Here, however, each document can belong to two or more (sometimes many) classes
at the same time.
The most commonly used way of dealing with the situation is to induce a
different classifier for each class. Whenever a future document’s class is needed,
the document’s attribute vector is submitted to all of these classifiers in parallel;
some of them will return “1” (meaning the document belongs to the corresponding
class), others will return a “0.”
This, however, makes the adventure even more expensive because of the need
to induce hundreds of classifiers. And given that the induction of each of these
classifiers is expensive in its own right, the result can be prohibitive.
Moreover, the classes are not independent of each other. At the very least, some
are subclasses or others, giving rise to a whole hierarchy of classes. This means that
their corresponding classifiers perhaps should not be induced in a manner totally
independent of the induction of the other classifiers.
• How are examples typically described in text-classification domains? Why are
the attribute vectors so long?
• What did this section mean by “graded class labels”?
• What makes induction of text classifiers expensive?
• Discuss the problem of multi-label examples.
• The number of examples that can be used for learning depends on the particular
application. In some domains, examples are abundant; for instance, if they can
be automatically extracted from a database. In others, they may be rare and
expensive, as was the case of the oil spill domain.
• We may be totally in the dark as to which attributes really matter. This was
the case of the oil spill domain where the shape and other characteristics could
be described by a virtually unlimited number of attributes obtained by image-
processing techniques.
• In some domains, the really important attributes are not known or cannot be
obtained at reasonable costs. Inevitably then, the induction has to use those
attributes that are available, even if the performance of the classifier thus induced
will be diminished.
• In the brain–computer-interface domain, it turned out to be enough that a majority
of the decisions were correct: in this event, the cursor landed in the intended
rectangle. We can see that the maximum classification accuracy may not be
strictly necessary.
• In domains with more than just two classes, the error rate does not give the
full picture of the classifier’s behavior. Quite often, some classes will be almost
always perfectly recognized while others will pose problems. It is therefore
important to know not only the average performance, but also the performance
on each individual class.
• In some domains, we want to be able to explain the decision. High classification
accuracy is not enough. This is the case of medical diagnosis.
• The costs of false positives can be different from the costs of false negatives. In
the oil-spill domain, it was difficult to express these costs in monetary terms.
• Many applications are interested in minimizing error rate. Sometimes, however,
error rate fails to give the full picture.
• Another important requirement: the possibility that the user be able to tweak
a parameter that will trade false positives for false negatives (see the oil spill
domain)
Historical Remarks The results from Table 8.1 are taken from Kononenko et
al. [48] who refer there to an older project of theirs. The oil-spill project was
reported in Kubat et al. [52]. The sleep classification task is addressed by Kubat
et al. [50], and the experience with using machine learning in brain–computer
interface was published by Kubat et al. [53]. The character-recognition problem
has been addressed for decades by the field of Computer Vision; the first major text
systematically addressing the issue from the machine-learning perspective seems to
be Mori et al. [69]. The text-classification task was first studied by Lewis and Gale
[55].
A minor remark is in place, here. Of these six applications, the author of this
book participated in two. His motivation for including his own work was not to stun
the reader with his incredible scholarship and inestimable contribution to the field
of machine learning. Far from it. The truth is, if you work on a project for quite
some time, you develop not only personal attachment to it, but also certain deeper
understanding that makes this particular application more appropriate for instruction
purposes than those which you only read about in the literature.
8.8 Exercises and Thought Experiments
The exercises are to solidify the acquired knowledge. The ambition of the suggested
thought experiments is to let the reader see this chapter’s ideas in a different light
and, somewhat immodestly, to provoke his or her independent thinking.
1. Suppose that you know that the correctness of some class labels in your training
set is not certain. Would you recommend that these “unreliable” examples be
removed from the training set? In your considerations, do not forget that some
of the pre-classified examples will be used as testing examples to assess the
classification performance of the induced classifier.
2. Discuss the reasons why application-domain users may be reluctant to accept
machine-learning tools. Suggest ways to eliminate, or at least diminish, their
suspicions and concerns.
3. Section 8.2 mentioned a simple mechanism by which the k-NN classifier can
manipulate the two kinds of error (false negatives versus false positives). Suggest
a mechanism that a Bayesian classifier or a neural network can use to the
same end. How would you go about implementing this mechanism in a linear
classifier?
4. In more than one of the domains discussed in this chapter, it was necessary
to reduce the number of irrelevant and/or redundant attributes. In the projects
reported here, decision trees were used. Suggest another possibility that would
use a technique from one of the previous chapters. Discuss its advantages and
disadvantages.
5. Suppose you wanted to implement a program that would decide whether a given
position in the tic-tac-toe game (see Fig. 8.7) is winning. What attributes would
you use? How would you collect the training examples? What can you say about
expected noise in the data thus collected? What classifier would you use? What
difficulties are to be expected?
Fig. 8.7 In tic-tac-toe, the

goal is to achieve three o x x o x x
crosses (or circles) in a row or x o x o
a column or on a diagonal
o x o
8.8 Exercises and Thought Experiments 171
1. Some realistic data sets for machine-learning experimentation can be found on

the website of the National Institute of Standards and Technology (NIST). Find
this web site, then experiment with some of these domains.
2. Go to the web and find a website about the demography of the 50 states in the
U.S. Identify an output variable whose value will be deemed positive if it is above
the U.S. average and negative otherwise. Each state thus constitutes an example.
Based on the information provided by the website, identify the attributes to
describe the examples. From the data thus obtained, induce a classifier to predict
the output variable’s value.
Chapter 9
Induction of Voting Assemblies
A popular way of dealing with difficult problems is to organize a brainstorming

session in which specialists from different fields share their knowledge, offering
diverse points of view that complement each other to the point where they may
inspire innovative solutions. Something similar can be done in machine learning,
too. A group of classifiers is created in a way that makes each of them somewhat
different. When they vote about the recommended class, their “collective wisdom”
often compensates for each individual’s imperfections.
This chapter deals with mechanisms for the induction of such sets of classifiers
from data. The reasons behind the high performance of these “classifier assemblies”
are explained on the simplest approach known as bagging. Building on these
foundations, we proceed to the more sophisticated boosting algorithms and their
variations, including the so-called stacking approach.
9.1 Bagging
For the sake of simplicity, we will limit ourselves to the field of two-class domains
where each example is either positive or negative. As always, the ultimate classifier
is to be induced from a set of pre-classified training examples.
The Underlying Principle The approach known under the name of bagging1
induces a group of classifiers. When presented with an example, the classifiers are
used in parallel, each offering an opinion as to which class the example should
be labeled with. A “master classifier” collects this information, and then chooses
the label that has received more votes.
1
The name is an acronym of booststrap aggregation.

DOI 10.1007/978-3-319-63913-0_9
174 9 Induction of Voting Assemblies
Table 9.1 The algorithm of Bagging

Input: the training set, T, and the user’s choice of the induction technique
1. Using random sampling with replacement, create from T several training subsets,
T1 ; T2 ; : : : Tn . Each subset consists of the same number of examples.
2. From each Ti , induce classifier Ci (for i D 1 : : : ; n).
3. When presented with a new example, submit it to all Ci ’s in parallel and let each classifier,
Ci , offer its recommendation for the example’s class.
4. A “master classifier” decides which of the two classes received more votes.
Assuming that each of the participating classifiers represents a somewhat

different aspect of the recognition task, the classifier group (sometimes called a
“voting assembly”) is expected to outperform any individual.
Induction of the Voting Assembly The principle of bagging is summarized in
Table 9.1. The idea is to take the original training set, T, and to create from it a
certain number of training subsets, T1 ; : : : Tn , each of the same size.
Once the subsets Ti have been created, a machine-learning algorithm induces a
classifier, Ci , separately from each of them. For this, any of the induction techniques
from the previous chapters can be used. However, the baseline version of bagging
assumes that the same technique (say, induction of decision trees with the same
user-set parameters) is always used.
Bootstrapping Let us now explain how the training examples for Ti are selected.
Each example has the same chance of being picked. Once it has been included in
Ti , it is “returned” to T, by which we mean that it will participate in the selection of
examples for TiC1 with the same probability as any other example. For a training set,
T, consisting of N examples, the selection is repeated n times in a process known as
bootstrapping.
An example can appear in Ti more than once and, conversely, some examples
will fail to appear in Ti . This means that each Ti consists of N examples (with
duplicities), but each of these training subsets is different. As a result, each of the
induced classifiers will focus on a different aspect of the learning problem.
Why It Works Figure 9.1 explains how the voting can help reduce the error
rate. Three classifiers are considered. If the errors are rare, there is a chance that
each classifier will err on different examples. In this event, each example will be
misclassified at most once, the other two class labels being correct. An individual’s
occasional mistake will thus be “outvoted” (and thus corrected) by the others.
Of course, the situation will rarely be so convenient. If an example is misclas-
sified by two out of the three classifiers, the voting will result in the wrong class.
One can stipulate, however, that this unfavorable situation might be improved if the
number of classifiers is increased.
9.1 Bagging 175
Fig. 9.1 The three classifiers

were asked to classify the
classifier 1:
same 17 examples. By doing classifier 2:
so each erred on three
examples—different for each classifier 3:
classifier. The reader can see
that these errors can be
eliminated by voting
... a correctly classified example
... an incorrectly classified example
The most important lesson to be drawn is that bagging works well if each
classifier tends to err on different examples.
Some Observations Practical experience shows that bagging achieves good results
if the error rate of each individual classifier is low. Then, with sufficiently large
number of classifiers, chances are high that an individual’s errors will be corrected
by the others, similarly as in Fig. 9.1.
This, however, is not necessarily the case when the error rates of the individual
classifiers are high. Up to a certain point, these errors are still corrected if there
are really a great many classifiers because then, by the law of large numbers, each
aspect of the underlying recognition problem is likely to be represented. The only
thing to criticize is that perhaps too many classifiers are then needed.
Conceptually, however, the approach involves too much randomness. Figure 9.1
leads us to believe that the same effect might be achieved if the classifiers were
induced not independently from each other, but rather in a way that might increase
their ability to complement each other. Such mechanisms will be the subject of
the following sections.
• What makes us believe that a group of voting classifiers will outperform an
individual? Under what circumstances will the scheme fail?
• How are the individual classifiers induced in the bagging approach?
• What is meant by bootstrapping?
9.2 Schapire’s Boosting
Although the bagging approach often achieves impressive results, it suffers from a
serious shortcoming: the voting classifiers have all been induced independently of
each other from randomly selected data. One would surmise that a smarter—and
perhaps more successful—approach should rely on a mechanism that makes the
classifiers complement each other. For instance, this can be done by inducing each
of them from training examples that are perceived as difficult by the other classifiers.
Schapire’s boosting was invented with this idea in mind.
Induction of Three Mutually Complementing Classifiers Suppose that a random
subset, T1 2 T, of m training examples has been created. These are used to induce
the first classifier, C1 . When testing this classifier on the entire training set, T, we
will observe that it misclassifies a certain number of examples.
Suppose we now create another training subset, T2 2 T. Let it consist of m
examples selected in a manner that ensures that the previously induced C1 classifies
correctly 50% of them, misclassifying the remaining 50%. This means that T2 is
so difficult for C1 that the classifier will not outperform a flipped coin. From the
training subset thus created, the second classifier, C2 , is induced.
The two classifiers, C1 and C2 , having been induced each from different
examples, will inevitably differ in how they label certain instances. A tie-breaker
is therefore needed. To this end, a third training subset, T3 , is created, consisting
only of examples on which C1 and C2 differ. From this third subset, T3 , the third
classifier, C3 , is induced.
The principle is summarized by the pseudocode in Table 9.2. When an example
is presented, the master classifier collects the labels recommended by the three
classifiers, and then returns the class that has received more votes.
Ideally, each of the training sets, Ti , is of the same size, m.
Recursive Implementation The principle just described can be implemented
recursively. Figure 9.2 shows how. The resulting triplet of classifiers (in the dotted
rectangle) is treated as a single classifier. In the next step, a new training subset, T4 ,
is created in a manner that ensures that the triplet’s error rate on T4 is 50%. From
Table 9.2 The algorithm of Schapire’s boosting

Input: the training set, T, and the user’s choice of the induction technique
1. Create a random training subset, T1 , and induce from it classifier C1 .

2. Create a training subset T2 in a manner that makes sure that C1 scores 50% on it. Induce
from T2 classifier C2 .
3. Create T3 such that C1 and C2 disagree on each of the examples it contains. Induce from
T3 classifier C3 .
4. For classification, use plain majority voting.
9.2 Schapire’s Boosting 177
master classif.B
master classif.A
classif.1 classif.2 classif.3 classif.4 classif.5
Fig. 9.2 Recursive application of Schapire’s boosting. Master classifier A combines the votes of
classifiers 1–3. Then, master classifier B combines the votes of master classifier A with those of
classifiers 4 and 5
these, classifier C4 is induced. Finally, training subset T5 is created from examples

on which the triplet and C4 differ; from these, classifier C5 is induced. Figure 9.2
also shows how the voting in the resulting scheme is hierarchically organized.
It is important to understand the logic of the voting procedure. If classifiers 1
through 3 all return 0, master classifier A returns 0, too; however, this result can still
be outvoted if both classifier 4 and classifier 5 return 1. The whole structure thus
may return 1 even if three out of the five participating classifiers return 0.
Note that the total number of classifiers induced using this single level of
recursion is 3 C 2 D 5. Of course, the principle can be repeated, resulting in
5 C 2 D 7 classifiers, and so on. Assuming NR levels of recursion, the total number
of participating classifiers is 2NR C 3.
Performance Considerations Suppose that each of the induced classifiers has an
error rate below a certain
. It has been proved that the voting triplet’s error rate is
less than 3
2 2
3 , which is always smaller than
. For instance, if
D 0:2, then
3
2 2
3 D 2 0:04 2 0:008 D 0:104. And if
D 0:1, then 3
2 2
3 D
0:03 0:002 D 0:028.
Put another way, Schapire’s boosting brings about an improvement over the
performance of the individual classifiers. But if the first voting triplet achieves (see
the previous paragraph) 3
2 2
3 D 0:104, it may be difficult to achieve an equally
low error rate with the other two classifiers (4 and 5). One possibility to handle this
situation is to create each of them (classifier 4 and classifier 5) as a triplet in its own
right as indicated in Fig. 9.3. In this case, the total number of classifiers participating
in NR levels of recursion will be 3NR .
The thing to remember is that each added level of recursion reduces the error
rate. It is therefore theoretically possible to reduce the error rate all the way down to
zero. Practice, however, is different—for reasons addressed in the next paragraph.
Limitations The main difficulty is how to find the right examples for each of the
subsequent training subsets. These, we already know, have to satisfy certain criteria.
master classif.B
master classif. master classif. master classif.
class.1 class.3 class.4 class.6 class.7 class.9

class.2 class.5 class.8
Fig. 9.3 Another approach to recursive application of Schapire’s boosting
Recall that, to induce the second classifier, C2 , we need to create a training subset,
T2 , that has been chosen in a way that makes the previous classifier, C1 , classify
correctly only 50% of them.
This may be easier said than done. For instance, if the entire training set that we
have at our disposal consists of 100 examples, and the first classifier’s error rate is
10%, then we have only 10 misclassified examples, to which 10 correctly classified
examples are added to create T2 ; this means that the size of T2 cannot exceed 20.
In the same spirit, we may find it impossible to find m examples that satisfy the
requirements for the third training subset, T3 , which, as we know, should consist
only of examples for which C1 and C2 give the opposite class labels.
And even if we do succeed in creating all the three training subsets, we may find
it impossible to apply the same principle recursively. To be able to do this, we would
need an almost inexhaustible (and certainly not affordable) source of examples—a
luxury rarely available.
To make sure you understand this topic, try to answer the following questions. If you
have problems, return to the corresponding place in the preceding text.
• How does Schapire’s boosting create the training subsets from which to induce
the individual classifiers?
9.3 Adaboost: Practical Version of Boosting 179
• Explain the two ways of implementing the principle recursively. How many
classifiers are induced in each case if NR levels of recursion are used? Explain
also the voting mechanism that reaches the final classification.
• When compared to the error rate of the individual classifiers, what is the error
rate of the final classification?
• Explain the practical limitation of Schapire’s boosting: the problem with finding
enough examples.
9.3 Adaboost: Practical Version of Boosting
The main weakness of bagging is the randomness of its behavior: each of the
constituting classifiers is induced in total isolation from all the other classifiers.
In Schapire’s boosting, the randomness is minimized, but we have noticed another
complaint: in a realistic setting, it is often impossible to find the right number of the
training examples that satisfy the conditions for the second and the third classifier.
And so perhaps the most practical approach to boosting is that of Adaboost where
the examples are chosen from a probabilistic distribution that is gradually modified
in a way that makes the whole “assembly” focus on those aspects that appear to be
difficult.
The General Approach Similarly to Schapire’s approach, Adaboost creates the
classifiers one by one, each of them from a different training subset whose
composition depends on the behavior of the previous classifiers. There is a major
difference, though. Whereas Schapire’s boosting selects the examples according to
certain precisely defined conditions, Adaboost chooses them probabilistically. This
means that each example has a certain probability of being drawn, the probabilities
of all examples summing up to 1. The mechanism is implemented in a way
that makes sure that examples that were repeatedly misclassified by the previous
classifiers will get a higher chance of being included in the training subset for the
induction of the next classifier.
Another difference is the number of classifiers. Unlike Schapire’s triplets, this
approach typically relies on a great number of classifiers. Further on, the final
decision is not achieved by the plain voting used in bagging. Instead, the so-called
weighted majority voting scheme is used (see later).
Probabilistic Selection of Training Examples We have mentioned that, for the
individual training subsets, Ti , the examples are selected from the original set, T,
probabilistically. Here is a simple way of implementing this principle. Suppose the
i-th example’s probability is p.xi / D 0:1. A random-number generator is asked for
a number between 0.0 and 1.0. If the returned number is from the interval Œ0:0; 0:1,
then the example is chosen; otherwise, it is not chosen.
At the beginning, when the first training set, T1 , is being created, each example
has the same chance of being selected. Put more technically, if T consists of m
examples, then the probability of each example is p D 1=m.
For each of the next subsets, Ti , the probabilities of the individual examples are
modified based on the observed behavior of the previous classifier, Ci1 . The idea
is to make sure that examples misclassified by the previous classifiers should get a
higher chance of being included in Ti . This will focus the next classifier, Ci , on those
aspects that the previous classifiers found difficult.
Modifying the Probabilities for the i-th Training Set To begin with, all m
training examples have the same probability, p D 1=m, of being selected. Using
this probability, the first training subset, T1 , is created. From T1 , the first classifier,
C1 is induced. After this, the probabilities of the training examples are modified
according to the classifier’s behavior. More specifically, the probability of examples
that have been correctly classified by C1 will be reduced, thus increasing the future
chances of examples misclassified by C1 .
The concrete way in which the probabilities are modified according to the i-
th classifier’s behavior is shown in Table 9.3. First, Adaboost calculates the i-th
classifier’s overall error,
i , as the weighted sum of errors on the whole original
training set: this is obtained simply by summing up the probabilities of the
misclassified examples. Once the weighted sum has been obtained, it is used to
reduce the probability of those examples that have been correctly classified by Ci .
To be more specific, each such probability is multiplied by the term, ˇi D
i =.1
i /.
Normalizing the Probabilities After this, the probabilities are normalized to make
sure that their values sum up to 1 as required by the theory of probability. The easiest
way to normalize them is to divide each probability by the sum of all probabilities.
For instance, suppose that the following probabilities have been obtained:
p1 D 0:4; p2 D 0:2; p3 D 0:2
Table 9.3 The algorithm of Adaboost

Input: the training set, T, consisting of m examples; and the user’s choice of the induction
technique
1. Let i D 1. For each xj 2 T, set p1 .xj / D 1=m.

2. Create subset Ti consisting of m examples randomly selected according to the given
probabilities. From Ti , induce Ci .
3. Evaluate Ci on each example, xj 2 T.
Let ei .xj / D 1 if Ci misclassified xj and ei .xj / D 0 otherwise.
Pm
i) Calculate
i D jD1 pi .xj /ei .xj /;
ii) Calculate ˇi D
i =.1
i /
4. Modify the probabilities of correctly classifiedPexamples by piC1 .xj / D pi .xj / ˇi
m
5. Normalize the probabilities to make sure that jD1 piC1 .xj / D 1.
6. Unless a termination criterion has been met, set i D i C 1 and go to 2.
9.3 Adaboost: Practical Version of Boosting 181
The sum of these values is 0:4 C 0:2 C 0:2 D 0:8. Dividing each of the three
probabilities by 0.8 will give the following normalized values:
0:4 0:2 0:2

p1 D D 0:5; p2 D D 0:25; p3 D D 0:25
0:8 0:8 0:8
It is easy to verify that the new probabilities now sum up to 1:
p1 C p2 C p3 D 0:5 C 0:25 C 0:25 D 1:0
An Illustration Table 9.4 shows how Adaboost modifies the probabilities of the
individual examples. At the beginning, each example has the same chance of being
selected for inclusion of the first training set, T1 . Once the first classifier, C1 , has
been induced from T1 , the classifier is applied to each single example from the
original set, T.
The probability of those examples that have been correctly classified by C1
P reduced according to piC1 .xj / D pi .xj / ˇi where
(examples x1 through x7 ) is then
ˇi D
i =.1
i / and
i D m jD1 pi .xj /ei .xj /. The resulting probabilities are then
normalized in order to make sure that they sum up to 1 as required by the theory of
probability.
After this, the second training set, T2 , is created, and then classifier C2 induced
from it. Based on the observed behavior of C2 when applied to T, we calculate the
values of
2 and ˇ2 ; these are then used when modifying the correctly classified
examples’ probabilities, p.x1 / : : : p.x10 /. The process thus indicated continues until
a predefined user-specified termination criterion is reached. This criterion can be
based on the maximum permitted number of classifiers having been reached, or
on the classification accuracy of the whole “assembly” having achieved a certain
threshold, or on some other requirement.
Weighted Majority Voting Once the classifiers have been induced, the example to
be classified is presented to them in parallel, and the final classification decision is
reached by a weighted majority voting whereby each classifier has been assigned
a certain weight according to its classification record (see below). The final
decision is not obtained by the plain vote used in the previous sections. Rather,
the voting mechanism used in Adaboost is somewhat less-then-democratic in that
each classifier has a different strength defined by its weight, denoted as wi .
When presented with an example, each classifier returns a class label. The final
decision is then obtained by comparing the sum, Wpos , of the weights of the
classifiers voting for the positive class with the sum, Wneg , of the weights of the
classifiers voting for the negative class.
For instance, suppose there are seven classifiers with the following weights:
Œ0:2; 0:1; 0:3; 0:7; 0:2; 0:9; 0:8. Suppose, further, that the first four of these return
pos, and last three return neg. Plain voting will return the pos label, seeing that
this class is supported by the greater number of votes.
Table 9.4 An example illustrating how Adaboost modifies the probabilities of the individual
examples
Suppose that the training set, T, consists of ten examples denoted as x1 : : : x10 . The total
number of examples being m D 10, all initial probabilities are set to p.xi / D 1=m D 0:1:
p1 .x1 / p1 .x2 / p1 .x3 / p1 .x4 / p1 .x5 / p1 .x6 / p1 .x7 / p1 .x8 / p1 .x9 / p1 .x10 /
0:1 0:1 0:1 0:1 0:1 0:1 0:1 0:1 0:1 0:1
According to this probability distribution, the examples for inclusion in T1 are selected.
From T1 , classifier C1 is induced.
Suppose that, when applied to T, classifier C1 classified correctly examples x1 : : : x7 (for
these, e1 .xj / D 0) and misclassified examples x8 : : : x10 (for these, e1 .xj / D 1).
The weighted error is then obtained as follows:

1 D †10
jD1 p1 .xj / e1 .xj / D 0:3
From here, the multiplicative term for probability modifications is calculated:
ˇ1 D
1 =.1
1 / D 0:43
The probabilities are modified by p.xj / D p.xj / ˇ1

Here are the new (not yet normalized) values:
0:043 0:043 0:043 0:043 0:043 0:043 0:043 0:1 0:1 0:1
After normalization, we obtain the following values:
0:07 0:07 0:07 0:07 0:07 0:07 0:07 0:17 0:17 0:17
Note that these ten probabilities sum up to 1.0. The next classifier, C2 , is then induced from a
training set T2 whose examples have been selected from T according to these last
probabilities.
By contrast, weighted majority voting will separately sum up the weights

supporting the positive class, obtaining Wpos D 0:2C0:1C0:3C0:7 D 1:3, and then
sum up the weights supporting the negative class: Wneg D 0:2 C 0:9 C 0:8 D 1:9.
The master classifier will then label the example as negative because, as we can see,
Wneg > Wpos .
How to Obtain the Weights of the Individual Classifiers Each classifier is
assigned a weight according to its performance: the higher the classifier’s reliability,
the higher its weight. On the other hand, the weight can in principle even be
negative—if it is believed that the classifier is more often wrong than right. It is
important to be aware of methods to find the concrete weights.
Many alternative possibilities exist. The inventors of Adaboost suggested specific
formulas which facilitated mathematical analysis of the technique’s behavior.
9.4 Variations on the Boosting Theme 183
Practically speaking, though, one can just as well use the perceptron-learning
algorithm from Chap. 4. The idea is to begin with equal weights for all classifiers,
and then present the system, one by one, with the training examples. Each time the
master classifier misclassifies an example, we increase or decrease the weights of
the individual classifiers according to the relation between the master classifier’s
hypothesis, h.x/, and the training example’s true class, c.x/.
Of course, other methods can be used. One might consider the WINNOW that
we know is good at weeding out those classifiers that do not contribute much to
the overall system’s performance (as if these classifiers were irrelevant attributes
describing an example).
• Describe the mechanism that Adaboost uses when selecting the training examples
to be included in the training set, Ti , from which the i-th classifier is to be induced.
• Explain how, after the induction of the i-th classifier, Adaboost modifies for each
example its probability that it will be chosen for inclusion in TiC1 .
• Explain the principle of the weighted majority voting that Adaboost uses when
deciding about a concrete example’s class label.
• How are the weights of the individual classifiers obtained?
9.4 Variations on the Boosting Theme
The essence of boosting is to combine several imperfect classifiers that tend to

complement each other. Schapire was the first to suggest a concrete way of inducing
the classifiers, with bagging and Adaboost being the most popular alternatives.
The number of variations on this theme is virtually inexhaustible. It is good to
be aware of them, and this is why this section presents at least some of the most
important ones.
Randomizing the Attribute Set Rather than inducing each classifier from a
different training subset (as in the previous versions of boosting), one may consider
using the same training examples, but always described by a different subset of
attributes.
Again, the input is the set, T, of the training examples—and the set, A, of the
attributes used to describe them. Instead of the random subsets of examples (as in
bagging), we choose N random subsets of attributes, Ai A. The i-th classifier
(i 2 Œ1; N) is then induced from the examples from T described by attributes from
Ai . As before, the classifiers’ outputs are combined by the weighted majority voting,
the weights being obtained (for instance) by perceptron learning.
The approach is useful in domains marked by a great many attributes of which
most are suspected to be either irrelevant or redundant. Typically, classifiers induced
using less valuable attribute sets will exhibit poor classification performance; and as
such, they will receive low (or even negative) weights.
The approach can easily be combined with classical bagging: the idea is that,
for each classifier, a different set of examples, and also a different set of attributes,
should be used.
Non-homogeneous Boosting So far, all boosting approaches presented here
assumed that the individual classifiers are induced from somewhat different data,
but always using the same induction technique. But there is no reason for this
always having to be the case. The so-called non-homogeneous boosting does the
exact opposite: each classifier is induced from the same data, but with a different
machine-learning technique. The classifiers’ outputs are then, again, combined by
weighted majority voting.
The main advantage of this approach is the way in which it reduces error. As we
will see in Chap. 10, the errors committed by any classifier fall into two fundamental
categories. The first is caused by the variance in the available data: from a different
training set, a somewhat different classifier will be induced, and this will lead to
different errors. The second source of error is the bias inherent in the classifier.
For instance, a linear classifier cannot avoid misclassifying some examples if the
decision surface separating the positive examples from the negative is highly non-
linear.
Importantly, non-homogeneous boosting is known to reduce both kinds of error:
variance-related errors (this reduction happens in all boosting algorithms) as well as
bias-related errors (which is an advantage specific to non-homogeneous boosting).
Stacking The idea of non-homogeneous boosting is to take the outputs of a group
of classifiers, and then reach the final classification decision by the weighted
majority voting. If you think of it, two layers are involved here: at the lower level are
the base-level classifiers; and at the upper level is the master classifier combining
their outputs. Note that the master classifier itself has to be induced; for instance,
with the help of perceptron learning—because this is essentially a linear classifier.
The principle can be generalized to the so-called stacking approach. As before, a
set of diverse classifiers is used at the lower lever. The upper level then goes beyond
the bounds of a linear classifier. Indeed, any paradigm can be used for the master
classifier: Bayesian, nearest-neighbor based, a decision tree, or a neural network.
The linear classifier used in non-homogeneous boosting is just one out of many
possibilities.
Likewise, the base-level classifiers may come from the most diverse paradigms.
Sometimes, however, the individual classifiers differ only in the concrete choice of
parameter values. For instance, there can be a few decision trees differing in the
extent of pruning. At any rate, the base-level classifiers should all differ in the way
they classify the examples.
9.5 Cost-Saving Benefits of the Approach 185
Table 9.5 The class labels x1 x2 x3 ... xm

suggested by the six
base-level classifiers are used Classifier 1 1 1 0 ... 0
as attributes to redescribe the Classifier 2 0 0 1 ... 1
examples Classifier 3 1 1 0 ... 1
Classifier 4 1 1 0 ... 1
Real class 1 1 0 ... 1
Each column then represents a training
example to be used for the induction of
the master classifier
The method is illustrated in Table 9.5. The rows represent the individual
classifiers. Specifically, six classifiers have been induced here, each by a different
machine-learning technique. The columns represent the training examples. The field
in the i-th row and the j-th column contains a one if the i-th classifier labels the
j-th example as positive; otherwise, it contains a zero. Each column can thus be
interpreted as a binary vector that redescribes the corresponding example. This new
training set is then presented to a machine-learning program that induces from it the
master classifier.
• Explain what this section meant by “randomizing the attribute set.” What is the
main advantage of the approach? How can it be combined with the classical
bagging?
• Explain the principle of non-homogeneous boosting. What is its main advantage
from the perspective of the errors committed by the classifier?
• Explain the two-layered principle of “stacking.” In what sense do we say that
stacking is a generalization of non-homogeneous boosting?
9.5 Cost-Saving Benefits of the Approach
In some machine-learning techniques, computational costs grow very fast with the
growing size of the training set. In this event, an experimenter will observe that
induction from half of the examples incurs only a small fraction of the costs incurred
by induction from all examples. Likewise, the induction technique’s costs may grow
very fast with the growing number of attributes.
Fig. 9.4 In some CPU time

machine-learning techniques,
computational costs grow T
quickly with the growing size
of the training set. Quite
often, induction from half of
the examples incurs only a
small fraction of the costs
incurred by induction from all
examples T/5
N/2 N #examples
An Illustration The situation is visualized in Fig. 9.4. Here, the time needed for
a hypothetical machine-learning technique to induce a classifier from N examples
is T; however, the time needed to induce a classifier from 50% of the examples is
only one fifth of it, 0:2T. From this follows that to induce two classifiers, one from
the first half of the training set, the other from the second half, we will need only
2 0:2T D 0:4T of the time that would have been needed for the induction of the
classifier from the whole set, T. The computational savings thus amount to 60% of
the original costs.
Generalization to K Classifiers Following the same logic, we may consider
induction of K classifiers, each from a different subset, Ti , such that the size of each
of them is m (a user-set constant) which is supposed to be much smaller than N.
The classifiers induced from these sets will then vote, just as they do in bagging. In
many cases, the induction of the individual classifiers will take only a tiny fraction of
the original time—and yet the classification performance of the “voting assembly”
thus obtained may compare favorably with that of a classifier induced from the
entire training set. Similar observations can be made also in the case of Schapire’s
boosting and Adaboost.
To conclude, the boosting paradigm may not only improve classification perfor-
mance. Quite often, it will lead to significant savings in the computational costs
associated with the induction. This observation is particularly important in domains
marked by many examples described by many attributes. Here, induction can be
very expensive, and any idea that helps reduce computational costs is more than
welcome.
Comments on Practical Implementation In bagging, the number of examples
to be included in Ti is the same as in the original training set; no computational
savings in the sense illustrated in Fig. 9.4 are thus possible. On the other hand, we
have to remember that the assumption about the sizes of Ti was used only to make it
possible to present bagging as a bootstrapping technique. In practical applications,
the assumption is unnecessary, and can thus be relaxed. In other words, the size, m,
of the sets Ti can be a user-set constant, just as it is in Adaboost.
In the case of Schapire’s Boosting and non-homogeneous boosting, the reader

will recall that the sizes of Ti , are user-specified parameters. The same applies to
stacking.
The Costs of Example Selection When considering the savings in computational
costs, however, we must not forget that the price for them is a certain unavoidable
overhead. To be more specific, additional costs are associated with the need to select
the examples to be included in the next Ti .
In the case of bagging, these costs are so small as to be easily neglected. They are
higher in Adaboost, but even here they are still affordable. The situation is different
in the case of Schapire’s boosting. Here, the search for the examples that satisfy the
conditions for T2 and T3 can be quite expensive—especially when a great number
of training examples have been made available, but only a small percentage of them
satisfy the required conditions.
• Elaborate on the statement that “computational costs of induction can be greatly
reduced by exploiting the idea of the voting assemblies induced from small
subsets of examples.”
• Discuss the circumstances under which one can expect that the boosting idea will
reduce computational costs.
• How expensive (computationally) is it to create the training subsets, Ti , in each
of the boosting techniques?
• A popular machine learning approach induces a set of classifiers, each from a

different subset of the training examples. When presented with an example, each
classifier offers a class. Based on these suggestions, a master classifier makes the
final decision.
• The simplest application of this idea is the so-called bagging. Here, the subsets
used for the induction of the individual classifiers are obtained from the original
training set, T, by a bootstrapping mechanism. Suppose that T contains m
examples. Then, when creating Ti , we choose m examples “with replacement.”
Some examples may then appear in Ti more than once, while others will not
appear there at all.
• Schapire’s boosting induces three classifiers, C1 ; C2 , and C3 , making sure that
they complement each other as much as possible. This complementarity is
achieved by the way the training subsets are created: the composition of T1
is random; the composition of T2 is such that C1 experiences 50% error rate
on this set; and T3 consists of examples on which C1 and C2 disagree. Each of
the three subsets contains the same number of examples.
• By contrast, Adaboost chooses the examples probabilistically in a way that makes
each Ti consist primarily of examples on which the previous classifiers failed.
Another difference is that the final class label is obtained by weighted majority
voting.
• A few variations on the theme exist. One of them, randomization of attribute
sets, induces the classifiers from the same training examples which, however,
are each time described by a different subset of the attributes. Another one,
non-homogeneous boosting, uses always the same training set (and the same
attributes), but induces each classifier by a different induction technique. The
final class is then obtained by weighted majority voting. Finally, stacking
resembles non-homogeneous boosting, but the output is decided by a master
classifier that has been trained on the outputs of the base-level classifiers.
Historical Remarks The idea of boosting was invented by Schapire [85] who
pointed out that, in this way, even the performance of very weak induction
paradigms can thus be “boosted”—hence the name. The somewhat more practical
idea underlying Adaboost was published by Freund and Schapire [28], whereas
the bagging approach was explored by Breiman [9]; its application to decision
trees, known under the term random forests, was published by Breiman [10]. The
principle of stacking (under the name Stacking Generalization) was introduced by
Wolpert [99].
Exercises
1. Suppose the probabilities of the training examples to be used by Adaboost are

those listed in Table 9.6. From these, a training subset, Ti , has been created, and
from Ti , classifier Ci was induced. Suppose that Ci then misclassifies examples
x2 and x9 . Show how the probabilities of all training examples are re-calculated,
and then normalize the probabilities.
2. Suppose that eight classifiers have labeled an example. Let the weights of the
classifiers returning the pos label be Œ0:1; 0:8; 0:2, and let the weights of the
Table 9.6 The probabilities p.x1 / p.x2 / p.x3 / p.x4 / p.x5 / p.x6 / p.x7 / p.x8 / p.x9 / p.x10 /
of ten training examples
0.07 0.07 0.07 0.07 0.07 0.07 0.07 0.17 0.17 0.17
classifiers returning the neg label be Œ0:1; 0:3; 0:3; 0:4; 0:9. What label is
going to be returned by a master classifier that relies on weighted majority voting?
3. Return to Table 9.5 that summarizes the class labels returned for some of the m
examples by six different classifiers. Suppose a 3-NN-based master classifier is
asked to label an example, and that the three nearest neighbors (the columns in
Table 9.5) are x1 , x2 , and x3 . What final label is returned?
1. Recall how Schapire’s boosting chooses the examples for inclusion in the training
sets T2 , and T3 . Discuss possible situations under which it is impossible to find
enough examples for these subsets. Identify a situation where enough examples
can be found, but the search for them is impractically expensive. Conversely,
under what circumstances can it be affordable to identify all the necessary
examples even when recursion is used?
2. Give some thought to the stacking approach. Think of a situation under which it
will disappoint. Conversely, suggest a situation where stacking will outperform
the less general non-homogeneous boosting.
1. Implement the basic algorithm of Adaboost. The number of voting classifiers is

determined by a user-set constant. Another user-set constant specifies the number
of examples in each training set, Ti . The weights of the individual classifiers are
obtained with the help of perceptron learning.
2. Apply the program implemented in the previous task to some of the benchmark
domains from the UCI repository.2 Make observations about this program’s
performance on different data. For each domain, plot a graph showing how
the overall accuracy of the resulting classifier depends on the number of
subclassifiers. Also, observe how the error rate on the training set and the error
rate on the testing set tend to converge with the growing number of classifiers.
3. Implement the stacking algorithm for different base-level learning algorithms and
for different types of the master classifier. Apply the implemented program to a
few benchmark domains, and observe its behavior.
2
Chapter 10
Some Practical Aspects to Know About
The engineer who wants to avoid disappointment has to be aware of certain

machine-learning aspects that, for the sake of clarity, our introduction to the basic
techniques had to neglect. To present some of the most important ones is the task
for this chapter.
The first thing to consider is bias: to be able to learn, the learner has to build on
certain assumptions about the problem at hand, thus reducing the size of the search
space. The next important point has to do with the observation that an increase in the
size of the training set can actually hurt the learner’s chances if most of the training
examples belong only to one class. After this, we will discuss the question how
to deal with classes whose definitions tend to change with context or in time. The
last part focusses on some more mundane aspects such as unknown attribute values,
the selection of the most useful sets of attributes, and the problem of multi-label
examples.
10.1 A Learner’s Bias
Chapter 7 boldly declared that there is “no learning without bias.” The point was that
an unconstrained (and hence extremely large) hypothesis space is bound to contain
many hypotheses that only by mere chance correctly classify the entire training set
while still erring a lot on future examples. There is another, more practical side to
it. To be able to find something, you need to know where to look; and the smaller
the place where that something is hidden, the higher the chances of finding it.
A Simple Example Suppose we are to identify the property shared by the follow-
ing set of integers: f2; 3; 10; 20; 12; 21; 22; 28g. In the language of machine learning,
these constitute positive examples. Alongside these, also negative examples are
provided: f1; 4; 5; 11g. In these, the property is not present.

DOI 10.1007/978-3-319-63913-0_10
192 10 Some Practical Aspects to Know About
A student trying to find the answer typically explores various notions offered by
the number theory, such as primes, odd numbers, integers whose values exceed a
certain threshold, results of arithmetic operations, and so on. After a lot of effort,
some property satisfying the training examples is found. Usually, however, the
discovered rule is ridiculously complicated and awkward to say the least.
And yet, there is a simple solution which the students almost never hit upon.
Thing is, the underlying property does not come from the realm of arithmetics. What
the positive examples have in common (and the negative examples lack) is that they
all begin with the letter t: two, three, : : :, all the way to twenty-eight. Conversely,
none of the integers in the set of negative examples begins with a t.
The reason this simple solution is so difficult to find is that most people search for
it in the wrong place: arithmetics. Expressed in the language of machine learning,
they rely on the wrong bias. Once they are given the correct answer, their mindset
will be willing to take this experience into account in the future. If you give them
another puzzle of a similar nature, they will subconsciously think not only about
arithmetics, but also about the English vocabulary—they will incorporate into their
thinking also this new bias.
Representational Bias Versus Procedural Bias As far as machine learning is
concerned, biases come in different forms. A so-called representational bias is
determined by the language in which we want the classifier to be formulated.
For instance, in domains with continuous-valued attributes, one possible repre-
sentational bias can consist in the choice of a linear classifier; another can be the
preference for polynomials, and yet another the preference for neural networks. If
all attributes are discrete-valued, the engineer may prefer conjunctions of attribute
values, or perhaps even decision trees. Of course, all of these biases then have its
advantages as well as shortcomings.
Apart from this, the engineer usually also has to opt for a certain procedural
bias. By this we mean preference to a certain method of searching for the solution,
the selection of a specific machine-learning procedure. For instance, one such bias
relies on the assumption that pruning will improve the classification performance of
a decision tree on future data. Another procedural bias is the choice of a concrete set
of parameters in a neural-network’s training. And yet another, in the field of linear
classifiers, can be the engineer’s decision to employ the perceptron learning instead
of WINNOW—or the other way round.
The Strength of a Bias Versus the Correctness of a Bias Suppose the engineer
wants to decide whether to approach the given machine-learning problem with a
linear classifier or with a neural network. If the positive and negative examples
are linearly separable, then the linear classifier is clearly the better choice. While
both paradigms are capable of finding the solution, neural networks tend to overfit
the training set, thus poorly generalizing to future data. On the other hand, if the
tentative boundary separating the two classes is highly non-linear, then the linear
classifier will lack the necessary flexibility, whereas neural networks will probably
manage quite easily. The reader now begins to understand that each bias has two
critical aspects: strength, and correctness.
10.1 A Learner’s Bias 193
A bias is strong if it defines only a narrow class of classifiers. In this sense, the
bias of linear classifiers is much stronger than that of neural networks: the former
only allow for linear decision surfaces, while the latter can model virtually any
decision surface.
A bias is correct if it is the right one for the task at hand. For instance, the linear
classifier’s bias is correct only in a domain where the positive examples are linearly
separable from the negative ones. A conjunction of boolean attributes is correct
only if the underlying class can indeed be described by a conjunction of attributes.
Of course, the opposite term, incorrect bias, is not a crisp concept. Some gradation
is involved; some biases are only slightly incorrect, others significantly so.
A Useful Rule of Thumb: Occam’s Razor Ideally, the engineer wants to use a
bias (representational or procedural) that is correct. And if there is a possibility to
choose between two or more biases that are all correct, the stronger bias is to be
preferred because it has a higher chance of success—this is what we learned in
Chap. 7 where the advice to choose the simpler solution was presented under the
name of the Occam’s Razor.
Unfortunately, we rarely know in advance the correctness/incorrectness of all
possible biases. An educated guess is the best we can hope for. In some paradigms,
say, high-order polynomials, the bias is so weak that there is a high probability that
a classifier from this class will classify the entire training set with zero error rate;
and yet its performance on future data is uncertain on account of its problems with
PAC-learnability. Strengthening the bias (say, by reducing a polynomial’s order)
will reduce the VC-dimension, increasing the chances on future data—but only as
long as the bias remains correct. At a certain moment, further strengthening of the
bias will do more harm than good because the bias becomes incorrect, perhaps very
much so.
What we need to remember is the existence of an almost inescapable trade-off: a
mildly incorrect but strong bias can be better than a correct but very weak bias. But
what the term, “mildly incorrect bias,” means in a concrete application can usually
be decided only based on the engineer’s experience or by additional experimentation
(see Chap. 11).
“Lifelong Learning” In some applications, the machine-learning software is to
learn a series of concepts or classes, all of which are expected to have a solution
within the realm of the same specific bias. In this event, it makes sense to organize
the learning procedure in two tiers. At the lower level, the task is to identify the most
appropriate bias; at the higher level, the software induces the classifier using this
bias. The term used for this strategy, “lifelong learning,” reminds us of something
typical of our own human difficulties in learning: the need to “learn how to learn”
in a given field.
Two Sources of the Classifier’s Errors The observations made so far will help us
get a better grasp of the two main sources of a classifier’s errors.
The first is the variance in the training examples. Thing is, the data used for
the induction of the concrete classifier almost never capture all aspects of the
underlying class. This is partly due to the way the training set has been created.
In some applications, the training set has been created at random. In other domains,
it consists of examples available at the given moment, which involves a great deal
of randomness, too. And in yet others, the training set has been created by an expert
who has chosen the examples which he believes best represent the given class. The
last case is inevitably subjective, and thus no less unreliable than the previous two. In
view of all this, one can easily imagine that a different training set might be created
for the same domain. And here is the point. From a different training set, a somewhat
different classifier will be induced, and this different classifier will make somewhat
different errors on future data. This is what we mean by saying that variance in the
training data is an important source of errors. Its negative effect can often be reduced
if we use really large training sets.
The second source of error is bias-related. If the two classes, pos and neg,
are not linearly separable, then any linear classifier is bound to misclassify certain
minimum percentage of future examples. Bias-related errors cannot be reduced
below a certain limit because they are inherent in the very nature of the selected
type of classifier.
It is instructive to give some thought to the trade-off between the two sources. For
one thing, the bias-related error can be reduced if we resort to a machine-learning
paradigm known to have a weaker bias. Unfortunately, one of the unintended
consequences of such a decision is higher variance. Conversely, variance can in
principle be reduced by strengthening the bias—which, if incorrect, will increase
the frequency of bias-related errors.
• Explain the difference between the representational bias and the procedural bias.
Illustrate each type by examples.
• Explain the difference between the strong and weak bias. Explain the difference
between the correct and incorrect bias. Discuss the interrelation of the two
dichotomies.
• What has this section taught us about the two typical causes of a classifier’s
underperformance on future data?
10.2 Imbalanced Training Sets
When discussing the oil-spill domain, Sect. 8.2 pointed out that well-documented
images of oil spills are relatively rare. Indeed, the project could rely only on a
few dozen positive examples while the negative examples were abundant. Such
10.2 Imbalanced Training Sets 195
imbalanced representation of the two classes is not without serious consequences

for machine learning. This section will explain the cause of the difficulties, then
proceed to some very simple ways of reducing their negative impact.
A Simple Experiment Suppose we have at our disposal a training set that is so
small that it consists of only 50 positive examples and 50 negative examples. Let
us subject this set to a fivefold crossvalidation1 : we divide it into five equally sized
parts; then, in five different experimental runs, we always remove one of the parts,
induce a classifier from the union of the remaining four, and test the classifier on
the removed part. In this manner, we eliminate, or at least reduce, the effect of
randomness in the choice of the concrete training set. At the end, we write down
the average results of the testing: classification accuracy on the positive examples,
classification accuracy on the negative examples, and the geometric mean of the two
classification accuracies.
Suppose now that we realize we have many more negative examples at our
disposal than we originally thought. In order to find out how this newly discovered
bounty is going affect the learning, we add to the previous training set another 50
negative examples (the positive examples remain the same), repeat the experimental
procedure, and then write down the new result. We then continue in the same spirit,
always adding to the training set another batch of 50 negative examples while
keeping the same original 50 positive examples.
Observation If we plot the results of the above series of experiments, we will
obtain a graph that, in all likelihood, will look very much like the one shown
in Fig. 10.1 where the 1-NN classifier was used. The reader can see that as the
number of the majority-class examples increases, the induced classifiers become
biased toward this class, gradually converging to a situation where the classification
accuracy on the negative examples (the majority class) approaches 100%, while the
classification accuracy on the positive examples (the minority class) drops to well
below 20%. The geometric mean of the two values keeps dropping, too.
The observation may appear somewhat counterintuitive. Surely the induced
classifiers should become more powerful if more training examples are made
available, even if these added examples all happen to belong to the same class? In
turns out, however, that the unexpected behavior described above is typical of many
machine-learning techniques. Engineers usually call it the problem of imbalanced
class representation.
Majority-Class Undersampling (The Mechanical Approach) The experiment
has convinced us that adding more examples from the majority class may cause
degradation of the induced classifier’s performance on the minority class. This may
be a serious shortcoming. Thus in the oil-spill domain, the minority class represents
the oil spills, the primary target of the machine-learning undertaking. In medical
diagnosis, any disorder we want to recognize is typically a minority class, too. And
1
An evaluation methodology introduced in Sect. 11.5.
1 nearest neighbor
100
90
80
70
accuracy
60
50
40
30
20
10
0 100 200 300 400 500 600 700 800
number of negative exambles
Fig. 10.1 Dotted curve: classification accuracy on the neg class; dashed curve: classification
accuracy on the pos class; solid curve: geometric means of the two classification accuracies
the same applies to software whose task is to alert a company to misuse of its product
(e.g., a wrongful use of calling cards, or credit-card fraud). In domains of this kind,
it is the minority class that interests us. We now know that blindly adding more
and more majority-class examples to the training set is likely to do more harm than
good.
Suppose we are provided with a heavily imbalanced training set where, say,
nine out of ten examples are negative. In this event, we will often benefit from the
removal of many negative examples. In the simplest possible approach, this removal
can be made at random: for instance, each negative example will face a 50% chance
of being deleted from the training set. As we noticed above, the classifier induced
from this reduced set is likely to outperform a classifier induced from the entire
training set.
Identifying the Cause The mechanical solution indicated in the previous para-
graph will hardly satisfy the thoughtful engineer who wants to understand why
the data-removing trick worked—or, conversely, why increasing the number of
majority-class examples may have such detrimental consequences.
Suppose the 1-NN classifier uses a training set where the vast majority of the
examples are negative, and only a few are positive. Moreover, the data are known
to suffer from a considerable amount of class-label noise. Limiting itself to an easy-
to-visualize two-class domain, the left part of Fig. 10.2 shows one such training set.
10.2 Imbalanced Training Sets 197
_ __ _ _ _ _ __ _ _
_ _ __
+ _ + __ _ + + _ _
+ _ _ + _ _
_+ _ + _
+ _+ _ +
+
_
_ _ _
Fig. 10.2 In noisy domains where negative examples heavily outnumber positive examples,
the removal of negative examples that participate in Tomek links may increase classification
performance
The reader is sure to have noticed the point: in consequence of the noise, the nearest
neighbor of almost every positive example is negative. In reality, these neighbors are
probably positive, their negative labels being explained by errors made in the course
of the creation of the training set. Be it as it may, the 1-NN classifier misclassifies
these positive examples, and this is why there are so many false negatives, and only
a few (if any) false positives.
Of course, not all machine-learning paradigms will suffer from this situation as
dramatically as the 1-NN classifier. But most of them will suffer to some degree,
and we now understand the reason why.
An Informed Solution: One-Sided Selection Knowing the source of our troubles,
we are ready to suggest a remedy. To wit, the cause of our woes is the presence
of many class-noisy examples in the positive region; the situation should therefore
improve if we remove primarily these examples (rather than resorting to a random
selection as in the mechanical approach suggested above).
In Chap. 3, we encountered a simple algorithm capable of identifying “suspi-
cious” examples: the technique of Tomek links. The reader will recall that two
examples, .x; y/, are said to participate in a Tomek link if three conditions are
satisfied: (1) each of the two examples has a different class label; (2) the nearest
neighbor of x is y; and (3) the nearest neighbor of y is x. In the situation depicted
in Fig. 10.2, many of the noisy examples on the left indeed do participate in Tomek
links. This indicates that we may do improve the classifier’s behavior if we delete
from the training set the negative participants of each Tomek-link pair. The principle
is known as one-sided selection because only one side of the Tomek link is selected
for inclusion in the training set.
Applying the technique to the training set shown in the left part of Fig. 10.2, we
will obtain something like the smaller training set shown in the right part. It is easy
to see that the frequency of false negatives is now going to be lower. The efficiency
of this methods is usually higher than just mechanical removal of randomly picked
negative examples.
The Opposite Solution: Oversampling the Minority Class In some domains,

however, the training set is so small that any further reduction of its size by
undersampling is impractical. Even the majority-class examples are sparse, here;
the deletion of any one of them may remove some critical aspect of the learning
task, and thus jeopardize the performance of the induced classifier.
In this event, the opposite approach is sometimes preferred. Rather than removing
majority-class examples, we add examples representing the minority class. Since we
do not have at our disposal real examples from this class, we have to create them
artificially. This can be done in two fundamental ways:
1. For each example from the minority class, create one copy and add this copy into
the training set. Alternatively, two or more copies for each example can thus be
created and added.
2. For each example from the minority class, create its slightly modified version and
add it into the training set. The modification is made by minor (random) changes
in continuous attributes; much less useful, though still possible, are changes in
discrete attribute values.
The left part of Fig. 10.2 helps us understand why this works. To the neighbor-
hood of some of the “afflicted” positive examples (those whose nearest neighbors
have been turned negative by noise), minority-class oversampling inserts additional
positive examples; as a result, the 1-NN classifier is no longer misled. The principle
helps improve the behavior of other types of classifiers, too.
To make sure you understand this topic, try to answer the following questions. If you
have problems, return to the corresponding place in the preceding text.
• What does the term imbalanced training set refer to? Explain the main reason
why induction from imbalanced training sets so often leads to disappointing
results.
• What is the essence of majority-class undersampling? Explain the mechanical
approach, and then proceed to the motivation and principle of the one-sided
selection that uses Tomek links.
• Explain the principle of minority-class oversampling. Describe and discuss the
two alternative ways of creating new examples that are to be added to the
training set.
10.3 Context-Dependent Domains 199
10.3 Context-Dependent Domains
Up till now, we have tacitly assumed that the underlying “meaning” of a given
class is fixed and immutable, that a single classifier, once induced, will under all
circumstances exhibit the same (or at least similar) behavior. This, however, is not
always the case.
Context-Dependent Classes Some classes change their essence with circum-
stances. If you think of that, this is the case of many concepts used in daily life.
Thus the meaning of “fashionable dress” changes in time, and different cultures
have a different idea of what they want to wear. “State-of-the-art technology” was
something else a 100 years ago that it is today. Even the intended meaning of such
notorious terms as “democracy” or “justice” depends on political background and
historical circumstances. And if you want a more technical example, consider the
problems encountered by speech-recognition software: everybody knows that the
same word is often pronounced differently in England than in North America; but
the software should “understand” speakers from both backgrounds.
Context-Dependent Features For the needs of this book, context is understood as
a “a feature that has no bearing on the class if taken in isolation, but still affects the
class when combined with other features.”
For instance, suppose you want to induce a classifier capable of suggesting
medical diagnosis, of recognizing X based on a set of symptoms. Some attributes,
say, gender, do not have any predictive power; the patient being male is no
proof of prostate-cancer; but the attribute value gender=female is a
clear indication that the class is not prostate-cancer. This, of course, was
an extreme sample. In other diagnoses, the impact of gender will be limited to
influencing the critical values of certain laboratory tests, say, p D 0:5 being a critical
threshold for male patients and p D 0:7 for female patients. Alternatively, the prior
probabilities will be affected, breast-cancer being more typical of females,
although men can suffer from it, too.
Induction in Context-Dependent Domains Suppose you want to induce a speech-
recognition system, and you have a set of training examples coming both from
British and American speakers. Suppose the attribute vector describing each
example contains the “context” attribute, the speaker’s origin. The other attributes
capture the properties of the concrete digital signal. Each class label represents a
different phoneme.
For the induction of a classifier that for each attribute vector decides which
phoneme it represents, the engineer can essentially follow two different strategies.
The first takes advantage of the contextual attribute, and divides the training
examples into two subsets, one for British speakers and one for American speakers;
then it induces a separate classifier from each of these training subsets. The second
strategy mixes all examples in one big training set, and induces one “universal”
classifier.
Practical experience indicates that, in applications of this kind, the first strategy
performs better, provided that the real-time system in which the classifiers are
embedded knows which of them to use. This decision can be assisted by an
additional two-valued classifier that is trained to distinguish British speakers from
American speakers.
Concept Drift Sometimes, the context changes in time. The “fashionable dress”
example mentioned earlier belongs to this category, as do the political terms. In this
event, machine-learning specialists talk about a so-called concept drift. What they
have in mind is that, in the course of time, the essence or meaning of a class drifts
from one context to another.
The drift has many aspects. One of them is the extent to which the context
has changed the meaning of the class. In some rare domains, this change is so
substantial that the induced classifier becomes virtually useless, and a new one has
to be induced. Much more typical, however, is a less severe change that results only
in a minor reduction of the classification performance. The old classifier can then
still be used, perhaps after some fine-tuning.
Another feature worth consideration is the “speed” of the drift. At one extreme
is an abrupt change. At a certain moment, one context is simply replaced, at it were,
by another. More typically, however, the change is gradual in the sense that there is
a certain “transition” period during which one context is, step by step, replaced by
the other. In this event, the engineer may ask how fast the transition is, and whether
(or when) the concept drift will necessitate special actions.
Induction of Time-Varying Classes Perhaps the simplest scenario in which
concept drift is encountered is the one shown in Fig. 10.3. Here, a classifier is faced
with a stream of examples that arrive one at a time, either in regular or irregular
intervals. Each time an example arrives, the classifier labels it with a class. There
may or may not be a feedback loop that tells the system (immediately or after some
delay) whether the classification was correct, and if not, what the correct class label
should have been.
If there is a reason to suspect the possibility of an occasional concept drift, it
may be a good idea to take advantage of a sliding window such as the one shown
in the picture. The classifier is then induced only from the examples “seen through
the window.” Each time a new example arrives, it is added to the window. Whenever
deemed appropriate, older examples are removed, either one at a time, or in groups,
Fig. 10.3 A window passes window

over a stream of examples; examples
“current classifier” is
periodically updated to reflect
changes in the underlying current
class. Occasionally, the classifier
system can retrieve some of store of
the previous classifiers if the old
underlying context recurs classifiers
10.3 Context-Dependent Domains 201
such as “the oldest 25% examples.” The motivation for the deletion is simple: the
engineer wants the window to contain only recent examples, suspecting that older
ones may belong to an outdated context.
As already mentioned, the classifier is supposed to reflect only the examples
contained in the window. In the simplest implementation, the classifier is re-induced
each time the window contents change. Alternatively, the change in the window
contents may only trigger a modification/adaptation of the classifier.
Figure 10.3 shows yet another aspect of this learning paradigm: sometimes, an
older context may reappear (e.g., due to a certain “seasonality”). For this reason, it
may be a good idea to store previously induced versions of the classifier, just in case
they might be re-used in the future.
Engineering Issues in the Sliding-Window Approach There are certain essential
issues that the engineer wishing to implement the sliding-window approach needs
to consider. The first is the question of the size of the window. If it is too small, then
the examples it contains may not be sufficient for successful learning. If it is too big,
then it may contain examples that come from outdated contexts. Ideally, then, the
window should grow (no old examples deleted) as long as it can be assumed that the
context has not changed. When a change is detected, a certain number of the oldest
examples should be deleted because they are no longer trusted.
This leads us to the next important question: how to recognize that a context has
changed? One simple solution relies on the feedback about the classifier’s behavior.
The change of context is then identified by a sudden drop in the classification
performance.
Finally, there is the question of how many of the oldest examples to delete from
the window. The answer will depend on how gradual the context change is, and also
on the extent of this change. At one extreme, an abrupt and considerable change will
call for the deletion of all examples. At the other extreme, a very slow transition
between two very similar contexts will necessitate the deletion of only a few of the
oldest examples.
• Give examples of domains where the meaning of a given class varies in
time and/or geographical location. Give examples of domains where previous
meanings recur in time.
• Describe the basic scenario that is based on a stream of time-ordered examples.
Explain the principle of the sliding-window approach to induction in time-
varying domains.
• Discuss briefly the basic engineering issues encountered in the sliding-window
approach.
10.4 Unknown Attribute Values
In many domains, certain attribute values are not known. A patient refused to
give his age, a measurement device failed, and some information got lost—or is
unavailable for any other reason. As a result, we get an imperfect training set such
as the one shown in Table 10.1 where some of the values are replaced with question
marks. In some domains, the question marks represent a considerable portion of
all attribute-value fields, and this may complicate the learning task. The engineer
needs to understand what kind of damage the unknown values may cause, and what
solutions exist.
Adverse Effects In the case of the plain version of the k-NN classifier, the distance
between two vectors can only be calculated if all values in the vectors are known.
True, the distance metric can be modified so that it quantifies also the distance
between, say, red and unknown; but distances calculated in this manner tend to
be rather ad hoc.
The situation is not any better in the case of linear and polynomial classifiers.
Without the knowledge of all attribute values, it is impossible to calculate the
weighted sum, †wi xi , whose sign tells the classifier which class label to choose.
Likewise, unknown attribute values complicate the use of Bayesian classifiers and
neural networks.
Decision trees are more flexible, in this sense. When classifying an example, it is
quite possible that the attribute whose value is unknown will not have to be tested
(will not find itself on the path from the root node to the terminal node).
Trivial Approaches to Filling-In Missing Values In a domain with a sufficiently
large training set that contains only a few question marks, there is usually no harm
in removing all examples that have unknown attribute values. This, however, will
become impractical in domains where the number of question marks is so high
Table 10.1 Training Crust Filling

examples with missing
Example Shape Size Shade Size Shade Weight Class
attribute values
ex1 Circle Thick Gray Thick Dark 7 pos
ex2 Circle Thick White Thick Dark 2 pos
ex3 Triangle Thick Dark Thick Gray 2 pos
ex4 Circle Thin White ? Dark 3 pos
ex5 Square Thick Dark ? White 4 pos
ex6 Circle Thick White Thin Dark ? pos
ex7 Circle Thick Gray Thick White 6 neg
ex8 Square Thick ? Thick Gray 5 neg
ex9 Triangle Thin Gray Thin Dark 5 neg
ex10 Circle Thick Dark Thick ? ? neg
ex11 Square Thick White Thick Dark 9 neg
ex12 Triangle Thick White Thick Gray 8 neg
10.4 Unknown Attribute Values 203
that the removal of all affected examples would destroy most of the training set,
disposing in the process of valuable information.
In this event, we may try to replace the question marks with some values, even
though these may be incorrect. This is easily done. When the attribute is discrete,
then we may simply replace the question mark with the attribute’s most frequent
value. Thus in Table 10.1, example ex8 , the unknown value of crust-shade will
be replaced with white because this is the most frequent value of this attribute in
this particular training set. In the case of a continuous-valued attribute, the average
value can be used. In ex6 and ex10 , the value of weight is unknown. Among
the 10 examples where it is known, the average value is weight=5.1, and this is
the value we will use in ex6 and ex10 .
When doing so, caution is called for. The reader has to keep in mind that using
the most frequent or average values will render the examples’ description unreliable,
perhaps even dubious. The technique should therefore be used sparingly. When
many values are missing, more sophisticated methods (such as the one below)
should be used.
Learning to Fill-in Missing Values Sometimes, using the most common or
average values can mislead the learning program. A better idea of how to fill the
empty slots is built around the observation that attributes are rarely independent
from each other. For instance, the taller the man, the greater his body weight. If the
weight of someone with height=6.5 is unknown, it would be foolish to use
the average weight calculated over the whole population; after all, our rather tall
individual is certainly heavier than the average person. Seeking a way out, we will
probably do better calculating the average weight among those with height > 6.
So much for a pair of mutually dependent attributes. Quite often, however,
the interrelations are more complicated than that, easily involving three or more
attributes. One simple mechanism to predict unknown values in situations of this
kind will rely on the idea of decision-tree induction. A pseudocode of the technique
is provided in Table 10.2.
The idea is quite simple. Suppose that at is an attribute that has, in the training
set, many question marks. We want to replace the question marks with the most
likely values. We decide to so by means of a decision tree. To this end, we convert
Table 10.2 An algorithm to determine unknown attribute values

Let T be the original training set.
Let at be the attribute with unknown values.
1. Create a new training set, T 0 , in which at becomes the class label; the examples are
described by all the remaining attributes, the former class label (e.g., pos versus neg)
being treated like just another attribute.
2. Remove from T 0 all examples in which the value of at is unknown.
3. From this final version of T 0 , induce a decision tree.
4. Use the decision tree thus induced to determine the values of at in those examples in
which its values were unknown.
the original training set, T, into a new training set, T 0 , where the original class label
(e.g., pos or neg) becomes one of the attributes, whereas at will be treated as a
class label. From this training set, we remove all examples whose values of at are
unknown. From the rest, we induce the decision tree, and then use the decision tree
to fill in the missing values.
• What are the main difficulties posed by unknown attribute values? What are their
typical consequences for classifier induction?
• Describe the trivial ways of dealing with examples with some unknown attribute
values. Discuss their limitations.
• Explain how the idea of decision-tree induction can be used when we want to
determine the unknown attribute values.
10.5 Attribute Selection
In many domains, the training examples are described by great many attributes:
tens of thousands, or even more. Learning from data sources of this kind can be
prohibitively expensive. Besides, we have to face problems with learnability, and
also with issues related to irrelevant or redundant attributes. To avoid unnecessary
disappointments, the engineer needs to be acquainted with methods to select the
most appropriate attributes.
Irrelevant and Redundant Attributes Not all attributes are created equal. Some
of them are irrelevant in the sense that their values do not have any effect on
an example’s class. Others are redundant in the sense that their values can be
obtained from values of other attributes: for instance, age can be obtained from the
value of date-of-birth. Attributes of these kinds can mislead certain induction
techniques. For instance, irrelevant attributes (and, to a lesser degree, also redundant
attributes) distort the vector-to-vector distances calculated by the k-NN classifier.
Other paradigms, such as decision trees, are less vulnerable, but even they may
suffer from excessive computational costs.
Extremely Long Attribute Vectors Some domains, such as automated text cat-
egorization, are marked by tens of thousands of attributes, and this often causes
problems. One of the difficulties is the reduced learnability: the induced classifiers
are prone to overfit the training data, disappointing the user when tested on future
examples. Also the computational costs can be impractical, especially when a
10.5 Attribute Selection 205
multilayer neural network is used: each additional attribute increases the number
of weights to be trained, thus adding to the calculations.
Moreover, examples described by thousands of attributes are inevitably sparse,
which is known to mislead many machine-learning approaches. For instance, the
problem of sparsity in k-NN classifiers was explained in Chap. 3.
And yet it is known that, for all intents and purposes, most of the attributes are
useless, and as such, should be disposed of.
Filter Approaches to Attribute Selection Perhaps the simplest approach to
attribute selection is based on what machine learning calls filtering. The idea is
to calculate for each attribute its “utility” for the classification task at hand, and then
order them according to this criterion. The intention to select the top N percent. The
choice of the “cut-off” point, N, is usually made by trial and error.
When ordering the attributes, the information gain from Sect. 6.3 can be used if
the attributes are discrete. Thanks to the existence of mechanisms for binarization
(see Sect. 6.4), information gain can actually be employed even in the case of
continuous-valued attributes; for this, however, statistical approaches to correlation
measurement are usually preferred.
One reason to criticize attribute filtering is that this approach ignores the relations
that exist among the attributes. This makes it difficult, almost impossible, to
identify redundant attributes. As we know, a redundant attribute does not bring
any additional information beyond that provided by the other attributes; and yet
its information gain can be high.
There is a relatively simple way to overcome this weakness. We induce a decision
tree, and then use only those attributes that are encountered in the tests in the tree’s
internal nodes. The careful reader will recall that this approach was used in some of
the simple applications discussed in Chap. 8.
Wrapper Approaches to Attribute Selection More powerful, but also more
computationally expensive, is the so-called wrapper approach to attribute selection.
Here is the underlying principle. Suppose we want to compare the quality of two
attribute sets, A1 and A2 . From the original training set, T, we create two training
sets, T1 and T2 . In both, all examples have the same class labels as in T. However, T1
describes the examples by A1 and T2 uses A2 . From the two newly created training
subsets, two classifiers are induced and evaluated on some independent evaluation
set, TE . The attribute set that results in the higher performance is better.
This is the principle used in the search-based algorithm whose pseudocode is
provided in Table 10.3. The input consists of a training set, T, and of a set of
attributes, A. The output is a subset, S 2 A, of the most useful attributes. At the
beginning, S is empty. At each step, the approach chooses the best attribute from A,
and adds it to S. What is “best” is determined by the classification performance (on
an independent testing set) of the classifier induced from the examples described
by the attributes from S. The algorithm is terminated if no addition to S leads to an
improvement of the classification performance—or if there are no more attributes to
be added to S.
Table 10.3 An wrapper approach to sequential attribute selection

Divide the available set of pre-classified examples into two parts, TT and TE . Let A be the set
of attributes. Create an empty set, S.
1. For every attribute, ati 2 A:
(i) add ati to S; let all examples in TT and TE be described by attributes from S;
(ii) induce a classifier from TT , then evaluate its performance on TE ; denote this
performance by pi ;
(iii) remove ati from S.
2. Identify the attribute that resulted in the highest value of pi . Remove this attribute from A
and add it to S.
3. If A D ;, stop; if the latest addition did not increase performance, remove this last attribute
from S and stop, too. In both cases, S is the final set of attributes.
4. Return to step 1.
• In what sense do we say that some attributes are less useful than others? Why
is the engineer often incapable of choosing the right attributes for example
description?
• Explain and discuss the principle of filter-based approaches to attribute selection.
• Describe the principle of the wrapper approaches to attribute selection. Explain
how the principle can be employed in a simple search-based approach to attribute
selection.
10.6 Miscellaneous
Some issues worth knowing about do not merit a separate section in an introductory
text, and yet they cannot be ignored. Let us briefly summarize them here.
Lack of Regularity in the Data Suppose you are asked to induce a classifier
from training examples that have been created by a random-number generator;
all attribute values are random, and so are the class labels. Obviously, there is no
regularity in such data—and yet—machine learning techniques are often capable of
inducing from them a classifier with zero error rate on the training set. Of course,
this perfect behavior on the training set will not translate into similar behavior on
future examples.
10.6 Miscellaneous 207
This observation suggests a simple mechanism to be used when measuring the

amount of regularity in data. The idea is simply to divide the data in two subsets—
one for training and one for testing. The classifier is induced from the training set,
and then applied to the testing set. In the case of random data, we will observe only
a small (if any) error rate on the training examples but the results on the testing
examples will be dismal. Conversely, the more regularity there is, in the data, the
better the results on the testing set.
Classes That Can Be Linearly Ordered In some domains, each example is
labeled with one out of several (perhaps even many) classes. In this context, we
have to mention the special case where the different class labels can be ordered. For
instance, suppose that the output class is month, with values january through
december. In domains of this kind, it would be a great (and misleading) simpli-
fication to assume that misclassifying june for may is the same as misclassifying
june for december.
Not only in performance evaluation, but also during the induction of such classes,
some attention should therefore be devoted to the ordering of the individual classes.
One possibility is to begin by grouping neighboring class labels; only after inducing
a classifier for each group should we consider the possibility of fine-tuning within
the group.
In the case of class month, we may perhaps first induce classifier for seasons,
each comprising 3 months, and only after this, a separate classifier for each month.
Regression In some applications, the expected output is not a discrete-valued class,
but rather a number from a continuous range. For instance, this can be the case when
the software is to predict a value of a stock-market index. Problems of this kind are
called regression. In this book, we do not address them. The simplest way of dealing
with them within the framework of machine learning is to replace the continuum
with subintervals, and then treat each subinterval as a separate class. Note that the
task would then belong to the category of “classes that can be linearly ordered”
mentioned in the previous paragraph.
• Explain how to use machine learning when measuring the amount of regularity in
data. Give examples of domains where this regularity can be expected to be low.
• What are multi-label domains? What is the simplest approach to induction
in these domains? What typical problems does the machine-learning engineer
encounter in them?
• Chapter 7 offered mathematical arguments supporting the claim that “there is no

learning without bias.” Certain practical considerations have convinced us that
this is indeed the case.
• Sometimes, the meaning of the underlying class depends on a concrete context;
this context can change in time, in which case we are facing the problem of time-
varying classes.
• The classical machine-learning techniques from the earlier chapters of this book
assume that both (or all) classes are adequately represented in the training set.
Quite often, however, this requirement is not satisfied, and the engineer has to
deal with the difficulties caused by the problem of imbalanced training sets.
• The most typical approaches to the problem of imbalanced training sets are
majority-class undersampling and minority-class oversampling.
• In many training sets, some attribute values are unknown, and this complicates
the use of certain induction techniques. One possible solution is to use (in place
of the unknown values) the most frequent or the average values of the given
attributes.
• Quite often, the engineer is faced with the necessity to select the most appropriate
set of attributes. Two fundamental approaches can be used here: the filtering
techniques and the wrapper techniques.
• In domains with more than two classes, it sometimes happens that the individual
classes can be ordered. This circumstance can affect performance evaluation. For
instance, if the task is to recognize a concrete month, then it is not the same thing
if the classifier’s output missed the target by 1 month or by 5. Even the learning
procedure should then perhaps be modified accordingly.
• Sometimes, the output is not a discrete-valued class, but rather a value from a
continuous range. This type of problem is called regression. This book does not
address regression explicitly.
Historical Remarks The idea to distinguish different biases in machine learning
was pioneered by Gordon and desJardin [34]. The principle of lifelong learning
was first mentioned by Thrun and Mitchell [88]. The early influential papers
addressing the issue of context were published by Turney [90] and Katz et al. [40].
Induction of time-varying concepts was introduced by Kubat [49] and some early
algorithms were described by Widmer and Kubat [97]. The oldest paper on multi-
label classification known to the author of this book was published by McCallum
[57]. The Wrapper approach to attribute selection is introduced by Kohavi [44].
Table 10.4 A simple Crust Filling

exercise in “unknown values”
ex3 Triangle Thick Dark ? Gray pos
ex4 Circle Thin White Thin ? pos
ex12 Triangle ? White Thick Gray neg
Exercises
1. Consider the training set shown in Table 10.4. How will you replace the missing
values (question marks) with the most frequent values? How will you use a
decision tree to this end?
2. Once you have replaced the question marks in Table 10.4 with concrete values,
identify the two attributes that offer the highest information gain.
1. The text emphasized the difference between two basic types of error: those
caused by the wrong bias(representational or procedural), and those that are due
to variance in the training data. Suggest an experimental procedure that would
give the engineer an idea about how much of the overall error rate can in a given
domain be explained by either of these two sources.
2. Boosting algorithms are known to be relatively robust with respect to variance-
based errors. Explain why this is the case. Further on, non-homogeneous boosting
presented in Sect. 9.4 is known to reduce bias-based errors. Again, offer some
explanation.
3. Suppose that a Bayesian classifier is to be employed in an imbalanced two-class
domain where examples from one class heavily outnumber examples from the
other class. Will this classifier be as sensitive to this situation as the nearest-
neighbor approach? Support your answer by concrete arguments. Suggest an
experimental verification.
4. In this section, the problem of imbalanced training sets was explored only within
the framework of two-class domains where each example is either positive or
negative. How does the same problem generalize to domains that have more than
two classes? Suggest some concrete situations where imbalanced classes in such
multi-class domains are or are not a problem.
5. Consider the case of linearly ordered classes mentioned in Sect. 10.6. Using the
hint provided in the text, suggest a machine-learning scenario addressing this
issue.
1. Write a computer program that accepts as input a training set where many
attribute-values are missing, and outputs an improved training set where the
missing values of discrete attributes have been replaced with the most frequent
values, and the missing values of continuous attributes with the average values.
Implement a computer program that will experimentally ascertain whether the
missing-values replacement helps or harms the performance of a decision tree
induced from such data.
2. Choose some public-domain data, for instance from the UCI repository.2 Make
sure this domain has at least one binary attribute. The exercise suggested here will
assume that this binary attribute represents a context. Divide the training data into
two subsets, each for a different context (a different value of the binary attribute).
Then induce from each subset the corresponding context-dependent classifier.
Assuming that it is at each time clear which of the two classifiers to use, how
much will the average performance of these two classifiers be better than that of
a “universal” classifier that has been induced from the original training set?
2
Chapter 11
Performance Evaluation
The previous chapters pretended that performance evaluation in machine learning

is a fairly straightforward matter. All it takes is to apply the induced classifier to a
set of examples whose classes are known, and then count the number of errors the
classifier has made. In reality, things are not as simple. Error rate rarely paints the
whole picture, and there are situations in which it can even be misleading. This is
why the conscientious engineer wants to be acquainted with other criteria to assess
the classifiers’ performance. This knowledge will enable her to choose the one that
is best in capturing the behavioral aspects of interest.
So much for the evaluation of classifiers. Somewhat different is the question
is how to compare the suitability of alternative machine-learning techniques for
induction in a given domain. Dividing the set of pre-classified examples randomly
into two subsets (one for induction, the other for testing) may not be the best thing to
do, especially if the training set is small; random division may then result in subsets
that do not represent the given domain properly. To obtain more reliable results,
repeated random runs are necessary.
This chapter addresses both issues, explaining alternative criteria to quantify the
performance of classifiers, and then discussing some strategies commonly used in
experimental evaluation of machine-learning algorithms. The question of statistical
evaluation of the results is relegated to the next chapter.
11.1 Basic Performance Criteria
Let us begin with formal definitions of error rate and classification accuracy. After
this, we will take a look at the consequences of the decision to refuse to classify an
example if the evidence favoring the winning class is weak.

DOI 10.1007/978-3-319-63913-0_11
212 11 Performance Evaluation
Table 11.1 The basic Labels

quantities used in the
returned by the classifier
definitions of performance
criteria pos neg
True labels: pos NTP NFN
neg NFP NTN
For instance, NFP is the number of false positives:
negative examples misclassified by the classifier
as positive
Correct and Incorrect Classification Let us first define four fundamental quan-
tities that will be used throughout this chapter. When testing a classifier on an
example whose real class is known, we can encounter only the following four
different outcomes: (1) the example is positive and the classifier correctly recognizes
it as such (true positive); (2) the example is negative and the classifier correctly
recognizes it as such (true negative); (3) the example is positive, but the classifier
labels it as negative (false negative); and (4) the example is negative, but the
classifier labels it as positive (false positive).
When applying the classifier to an entire set of examples (whose real classes are
known), each of these four outcomes will occur a different number of times—and
these numbers are then employed in the performance criteria defined below. The
symbols representing the four outcomes are summarized in Table 11.1. Specifically,
NTP is the number of true positives, NTN is the number of true negatives, NFP is the
number of false positives, and NFN is the number of false negatives. In the entire
example set, T, only these four categories are possible; therefore, the size of the set,
jTj, equals the sum, jTj D NFP C NFN C NTP C NTN .
Note that the number of correct classifications is the number of true positives plus
the number of true negatives, NTP C NTN ; and the number of errors is the number of
false positives plus the number of false negatives, NFP C NFN .
Error Rate and Classification Accuracy A classifier’s error rate, E, is the
frequency of errors made by the classifier over a given set of examples. It is
calculated by dividing the number of errors, NFP C NFN , by the total number of
examples, NTP C NTN C NFP C NFN .
NFP C NFN
ED (11.1)
NFP C NFN C NTP C NTN
Sometimes, the engineer prefers to work with the opposite quantity, classification
accuracy, Acc: the frequency of correct classifications made by the classifier over a
given set of examples. Classification accuracy is calculated by dividing the number
of correct classifications, NTP C NTN , by the total number of examples. Note that
Acc D 1 E.
NTP C NTN
Acc D (11.2)
NFP C NFN C NTP C NTN
11.1 Basic Performance Criteria 213
Rejecting an Example When discussing the problem of optical character recog-

nition, Sect. 8.1, suggested that the classifier should sometimes be allowed to
refuse to classify an example if the evidence supporting the winning class is
not strong enough. The motivation is quite simple: in some domains, the penalty
for misclassification can be much higher than the penalty for not making any
classification at all.
An illustrative example is not difficult to find. Thus the consequence of a
classifier’s refusal to return the precise value of the ZIP code is that the decision
where the letter should be sent will have to be made by a human operator. To be
sure, this manual processing is more expensive than automatic processing, but not
excessively so. On the other hand, an incorrect value returned by the classifier results
in the letter being sent to a wrong destination, which can cause a serious delay in
delivery. This latter cost is often much higher than the cost of “manual” reading.
Similarly, an incorrect medical diagnosis is often more expensive than no diagnosis
at all; lack of knowledge can be remedied by additional tests, but a wrong diagnosis
may result in choosing a treatment that does more harm than good.
This is why the classifier should sometimes refuse to classify an example if the
evidence favoring either class is insufficient. In some machine-learning paradigms,
the term insufficient evidence is easy to define. Suppose, for instance, that, in a 7-NN
classifier, four neighbors favor the pos class, and the remaining three favor the neg
class. The final count being four versus three, the situation seems “too close to call.”
More generally, the engineer may define a threshold for the minimum difference
between the number of votes favoring the winning class and the number of votes
favoring the runner-up class.
In bayesian classifiers, the technique is easily implemented, too. If the difference
between the probabilities of the two most strongly supported classes falls short of
a user-specified minimum, the example is rejected as too ambiguous to classify.
Something similar can be done also in neural networks: compare the signals returned
by the corresponding output neurons—and refuse to classify if there is no clear-cut
winner.
In other classifiers, such as decision trees, implementation of the rejection mech-
anism is not so straightforward, and is only made possible by the implementation of
“additional tricks.”
Advantages and Disadvantages of a Rejection to Classify The classifier that
occasionally refuses to make a decision about an example’s class is of course less
likely to go wrong. No wonder that its error rate will be lower. Indeed, the more
examples are rejected, the lower the error rate. But the caution should not be
exaggerated. It may look like a good thing that the error rate is reduced almost
to zero. But if this low rate is achieved only thanks to the refusal to classify almost
all examples, the classifier becomes impractical. Which of these two aspects (low
error rate versus rare classifications) plays a more important role will depend on the
concrete circumstances of the given application.
Figure 11.1 graphically illustrates the essence of the trade-off involved in
decisions of this kind. The horizontal axis represents a parameter capable of
error rate rejection rate
100%
Fig. 11.1 Error rate can be increased by allowing the classifier to refuse to classify an example if
available evidence fails to “convince” it. At the extreme, the error rate drops to E D 0 because all
examples are rejected
adjusting the rejection rate. As we move from left to right, the rejection rate
increases, whereas the error rate goes down until, at the extreme, all examples
are rejected. At this point, the zero error rate is a poor consolation for having
a classifier that never does anything. The lesson is clear. Occasional rejection of
unclear examples makes a lot of sense, but the principle should be handled with
care.
• Define the terms false negative, false positive, true negative, and true positive.
• Specify the formulas for classification accuracy and error rate. How are the two
criteria mutually interrelated?
• When should a classifier reject an example (i.e., when should it refuse to
classify)? How would you implement this behavior in diverse types of classifiers?
• How does rejection rate relate to error rate? What is the trade-off between the
two? Under what circumstances can refusal to classify be useful, and under what
circumstances does it hurt?
11.2 Precision and Recall
In some applications, negative examples outnumber positive ones by a wide

margin. When this happens, error rate offers a misleading picture of classification
performance. To see why, just consider the case where only 2% of all examples
11.2 Precision and Recall 215
are positive, and all the remaining 98% are negative. A “classifier” that returns the
negative class for any example in the set will be correct 98% of the time—which
may look like a remarkable feat. And yet, the reader will agree, a classifier that never
recognizes a positive example is useless.
Imbalanced Classes Revisited This observation is worth keeping in mind because
domains with imbalanced classes are quite common. We encountered some of
them in Chaps. 8 and 10, and other applications can be found. Thus in automated
information retrieval, the user may want to find a scientific document dealing with,
say, “performance evaluation of classifiers.” No matter how attractive the topic may
appear to this particular person, papers dealing with it will represent only a small
fraction of the millions of documents available in the digital library. Likewise,
patients suffering from a specific medical disorder are in the entire population
relatively rare. And the same goes for any undertaking that seeks to recognize a rare
event such as a default on mortgage payments or a fraudulent use of a credit card.
A seasoned engineer will go so far as to say that the majority of realistic applications
are in some degree marked by the phenomenon of imbalanced classes.
In domains of this kind, error rate and classification accuracy will hardly tell us
anything reasonable about the classifier’s practical utility. Rather than averaging the
performance over both (or all) classes, we need criteria that focus on a class which,
while important, is represented by only a few examples. Let us take a quick look at
some of them.
Precision By this we mean the percentage of true positives, NTP , among all
examples that the classifier has labeled as positive: NTP C NFP . The value is thus
obtained by the following formula:
NTP
Pr D (11.3)
NTP C NFP
Put another way, precision is the probability that the classifier is right when labeling
an example as positive.
Recall By this we mean the probability that a positive example will be correctly
recognized as such (by the classifier). The value is therefore obtained by dividing the
number of true positives, NTP , by the number of positives in the given set: NTP CNFN .
Here is the formula:
NTP
Re D (11.4)
NTP C NFN
Note that the last two formulas differ only in the denominator. This makes sense.
Whereas precision is the frequency of true positives among all examples deemed
positive by the classifier, recall is the frequency of the same true positives among
all positive examples in the set.
Table 11.2 Illustration of the two criteria: precision and recall

Suppose a classifier has been induced. Evaluation on a testing set gave the results summarized
in this table:
Labels
pos neg
True labels: pos 20 50
neg 30 900
From here, the following values of precision, recall, and accuracy are obtained:
20 20 920
precision D D 0:40I recall D D 0:29I accuracy D D 0:92
50 70 1000
Suppose the classifier’s parameters were modified with the intention to improve its behavior
on positive examples. After the modification, evaluation on a testing set gave the results
summarized in the table below.
Labels
pos neg
neg 20 880
From here, the following values of precision, recall, and accuracy are obtained.
30 30 910
precision D D 0:60I recall D D 0:30I accuracy D D 0:91
50 100 1000
The reader can see that precision has considerably improved, while recall remained virtually
unaffected. Note that classification accuracy has not improved, either.
Illustration of the Two Criteria Table 11.2 illustrates the behavior of the two
criteria in a simple domain with an imbalanced representation of two classes,
pos and neg. The induced classifier, while exhibiting an impressive classification
accuracy, suffers from poor precision and recall. Specifically, precision of Pr D
0:40 means that of the 50 examples labeled as positive by the classifier, only 20
are indeed positive, the remaining 30 being nothing but false positives. With recall,
things are even worse: out of the 70 positive examples in the testing set, only 20 are
correctly identified as such by the classifier.
Suppose that the engineer decides to improve the situation by modifying some of
the classifier’s internal parameters, and suppose that this modification results in an
increased number of true positives (from NTP D 20 to NTP D 30) and also a drop
in the number of false positives (from NFP D 30 to NFP D 20). On the other hand,
the number of false negatives has gone up, too: from NFN D 50 to NFN D 70. The
calculations in Table 11.2 indicate that recall was thus barely affected, but precision
11.2 Precision and Recall 217
has improved, from Pr D 0:40 to Pr D 0:60. Classification accuracy remained

virtually unchanged—actually, it has even gone down a bit, in spite of the improved
precision.
When High Precision Matters In some domains, precision is more important than
recall. For instance, when you purchase something from an e-commerce web site,
their recommender system often reacts with a message to the effect that, “Customers
who have bought X buy also Y.” The obvious intention is to cajole you into buying
Y as well.
Recommender systems are sometimes created with the help of machine learning
techniques applied to the company’s historical data.1 When evaluating their perfor-
mance, the engineer wants to achieve high precision. The customers better be happy
about the recommended merchandise, or else they will ignore the recommendations
in the future.
The value of recall is here unimportant. The list offered on the web site has to
be of limited size, and so it does not matter much that the system identifies only a
small percentage of all items that the customers may like.
When High Recall Matters In other domains, by contrast, recall is more impor-
tant. This is often the case in medical diagnosis. A patient suffering from X, and
properly diagnosed as such, represents a true positive. A patient suffering from X
but not diagnosed as such represents a false negative, something the doctor wants
to avoid—which means that NFN should be small. In the definition of recall, Re D
NTP
NTP CNFN
, the number of false negatives appears in the denominator; consequently, a
small value of NFN implies a high value of recall.
ROC Curves In many classifiers, tweaking certain parameters can modify the
values of NFP and NFN , thus affecting (at least to some extent) the classifier’s
behavior, for instance, by improving recall at the cost of worsened precision or
vice versa. This possibility can be useful in domains where the user has an idea as
to which of these two quantities is more important.
The reader will find it easy to suggest various ways to do the “tweaking.” Thus in
the k-NN classifier, the engineer may request that an example be labeled as negative
unless some very strong evidence supports the positive label. For instance, if four out
of seven nearest neighbors are positive, the classifier can be instructed still to return
the negative label (in spite of the small majority recommending the positive class).
In this way, the number of false positives can be reduced, though this often means to
the price in terms of an increased number of false negatives. Even stronger reduction
in NFP (and an increase in NFN ) can be achieved by requesting that any example
be deemed negative unless at least five (or six) of the seven nearest neighbors are
positive.
1
The concrete techniques employed to this end are somewhat more advanced than those discussed
in this textbook, and are thus not treated here.
Fig. 11.2 Example of ROC accuracy on positives

curves for two classifiers, c1
and c2. The parameters of the 1
classifiers can be used to
modify the numbers of false
positives and false negatives
c1
c2
0
0 error rate on negatives 1
Something similar is easy to implement in Bayesian classifiers and in neural

networks. The idea is the same: label the example with the preferred class unless
strong evidence indicates that this decision is incorrect.
The behavior of the classifier under different values of its parameters can be
visualized by the so-called ROC curve, a graph where the horizontal axis represents
error rate on the negative examples, and the vertical axis represents classification
accuracy on the positive examples. Example ROC curves for two classifiers, c1 and
c2, are shown in Fig. 11.2. Ideally, we would like to reach the upper-left corner that
represents zero error rate on the negatives and 100% accuracy on the positives. For
various reasons, this is rarely possible.
An important question to ask is which of the two, c1 or c2 is better. This can
only be answered based on the specific needs of the concrete application. All we
can say by just looking at the graph is that c1 outperforms c2 on the positives in the
region with low error rate on negatives. As the error rate on the negative examples
increases, c2 outperforms c1 on the positive examples. Again, whether this is good
or bad can only be decided by the user. See the comments concerning the question
when to prefer precision and when recall.
• In what kind of data would we rather evaluate the classifier’s performance by the
quantities known as precision and recall (instead of, say, error rate)?
• What formulas help us calculate the concrete values of these criteria? How do
the two formulas differ?
• Under what circumstances do we prefer high precision, and under what circum-
stances do we place more emphasis on high recall?
11.3 Other Ways to Measure Performance 219
• Explain the nature of an ROC curve. What extra information does the curve
convey about a classifier’s behavior? How does the ROC curve help the user
in the choice between two alternative classifiers?
11.3 Other Ways to Measure Performance
Apart from error rate, classification accuracy, precision, and recall, other criteria
are sometimes used, each reflecting a somewhat different aspect of the classifier’s
behavior. Let us take at least a cursory look at some of the most important ones.
Combining Precision and Recall in Fˇ Using two different (and sometimes
contradictory) performance criteria can be awkward; it is in the human nature to
want to quantify things by a single number. Especially in the case of precision and
recall, attempts have been made somehow to combine the two in one quantity. The
best-known such formula, Fˇ , is defined as follows:
.ˇ 2 C 1/ Pr Re
Fˇ D (11.5)
ˇ 2 Pr C Re
The parameter, ˇ 2 Œ0; 1/, enables the user to weigh the relative importance of
the two criteria. If ˇ > 1, then more weight is given to recall. If ˇ < 1, then more
weight is apportioned to precision. It would be easy to show that Fˇ converges to
recall when ˇ ! 1, and to precision when ˇ D 0.
Quite often, the engineer does not really know which of the two, precision or
recall, is more important, and by how much. In that event, she prefers to work with
the neutral value of the parameter, ˇ D 1:
2 Pr Re
F1 D (11.6)
Pr C Re
A Numeric Example Suppose that an evaluation of a classifier on a testing set

resulted in the values summarized in the upper part of Table 11.2. For these, the
table established the values of precision and recall, respectively, as Pr D 0:40 and
Re D 0:29. Using these numbers, we will calculate Fˇ for the following concrete
settings of the parameter: ˇ D 0:2; ˇ D 1, and ˇ D 5.
.0:22 C 1/ 0:4 0:29 0:121

F0:2 D 2
D D 0:39
0:2 0:4 C 0:29 0:306
.12 C 1/ 0:4 0:29 0:232
F1 D D D 0:70
0:4 C 0:29 0:330
.52 C 1/ 0:4 0:29 3:02
F5 D D D 0:29
52 0:4 C 0:29 10:29
Sensitivity and Specificity The choice of the concrete criterion is often influenced
by the given application field—with its specific needs and deep-rooted traditions
that should not be ignored. Thus the medical domain has become accustomed to
assessing performance of their “classifiers” (not necessarily developed by machine
learning) by sensitivity and specificity. In essence, these quantities are nothing but
recall measured on the positive and negative examples, respectively. Let us be more
concrete:
Sensitivity is recall measured on positive examples:
NTP
Se D (11.7)
NTP C NFN
Specificity is recall measured on negative examples:
NTN
Sp D (11.8)
NTN C NFP
In the machine-learning literature, evaluation in terms of sensitivity and speci-

ficity is rare, but it is still good to be aware of these two criteria. After all, we may
be asked to use them when applying machine learning to medical data—and quite
often it is the customer who is the ultimate boss.
Gmean When inducing a classifier in a domain with imbalanced class representa-
tion, the engineer sometimes wants to achieve similar performance on both classes,
pos and neg. In this event, the geometric mean, gmean, of the two classification
accuracies (on the positive examples and on the negative examples) is used:
s
p NTP NTN
gmean D accpos accneg D (11.9)
NTP C NFN NTN C NFP
Note that gmean is actually the square root of the product of two numbers:
recall on positive examples and recall on negative examples—or, in other words,
the product of sensitivity and specificity.
Perhaps the most important aspect of gmean is that it depends not only on the
concrete values of the two terms under the square root symbol, accpos and accneg ,
but also on how close the two values are to each other. A simple numeric example
will convince us that this is indeed the case.
Thus the arithmetic average of 0:75 and 0:75 is .0:75 C 0:75/=2 D 0=75; also
the arithmetic average of 0:55 and 0:95p is .0:55 C 0:95/=2 D 0:75. However, the
geometric means of the firstp pair is 0:75 0:75 D 0:75 whereas the geometric
means of the second pair is 0:55 0:95 D 0:72, a smaller number. We can see
that the gmean is indeed smaller when the two numbers are different; and the more
different they are, the lower the value of gmean.
11.3 Other Ways to Measure Performance 221
Another Numeric Example Again, suppose that the evaluation of a classifier on

a testing set resulted in the values summarized in the upper part of Table 11.2. The
values of sensitivity, specificity, and gmean are calculated as follows:
20
Se D D 0:29
50 C 20
900
Sp D D 0:97
900 C 30
r
20 900 p
gmean D D 0:29 0:97 D 0:53
50 C 20 900 C 30
Cost Functions In the world of machine-learning applications, not all errors carry
the same penalty. False positives may be more costly than false negatives or vice
versa. Comparisons of alternative classifiers are further complicated by the fact that,
in some domains, the costs of the two types of error cannot even be expressed in the
same (or at least comparable) units.
Consider the oil spill-recognition problem discussed in Chap. 8. A false positive
here means that a “lookalike” is incorrectly taken for an oil spill (false positive).
When this happens, a plane is unnecessarily sent to the given location to verify
the case. The costs incurred by this error are those associated with the flight. By
contrast, a false negative means that a potential environmental hazard has gone
undetected, something difficult to express in monetary terms.
• Explain how Fˇ combines precision and recall. Write down the formula and
discuss how different values of ˇ give more weight either to precision or recall.
Which value of ˇ gives equal weight to both?
• Give the formulas for the calculation of sensitivity and specificity. Explain their
nature, and show their relation to recall. When are the two criteria used?
• Give the formula for the calculation of gmean. Explain the nature of this quantity,
and show its relation to recall. When is this criterion used?
• Under what circumstances can the costs associated with false positives be
different from the costs associated with false negatives? Suggest a situation where
the comparison of the costs is almost impossible.
11.4 Learning Curves and Computational Costs
The first four sections of this chapter dealt with the problem of performance
evaluation of the induced classifiers. Let us now turn our attention to the evaluation
of the learning algorithm itself. How efficient is the given induction technique
computationally? And how good are the classifiers it induces? Will better results
be achieved if we choose some other machine-learning framework?
In this section, we will give some thought to the costs of learning—in terms of
the number of examples needed for successful induction, as well as in terms of the
computational time consumed. The other aspect, the ability to induce a tool with
high classification performance will be addressed in the following section.
The Learning Curve When evaluating a human subject’s ability to learn how
to solve a certain problem, psychologists rely on a learning curve, a notion that
machine learning has borrowed for its own purposes.
From our perspective, the learning curve simply shows how the classification
performance of the induced classifier depends on the size of the training set. Two
such curves are shown in Fig. 11.3. The horizontal axis represents the number
of training examples; and the vertical axis represents the classification accuracy
of the classifier induced from these examples. Usually, though not always, this
classification accuracy is evaluated on independent testing examples.
Most of the time, a larger training set means higher classification performance—
at least until the moment when no further improvement is possible. Ideally, we
would like to achieve maximum performance from the smallest possible training
set. For one thing, training examples can be expensive to obtain, and their source
can be limited no matter how much we are willing to spend on them. For another,
the more examples we use, the higher the computational costs on induction.
Comparing Learners with Different Learning Curves Figure 11.3 shows the
learning curves of two learners, l1 and l2 . The reader can see that the learning
curve of the former, l1 , rises very quickly, only to level off at a point beyond
which virtually no improvement is possible—the limitation may be imposed by an
Fig. 11.3 A learning curve classification accuracy

shows how the achieved
classification accuracy 100%
depends on the number of
examples used during
learning
l1
l2
# training examples
11.4 Learning Curves and Computational Costs 223
incorrect bias (see Sect. 10.1). By contrast, the learning curve of the second learner,
l2 , does not grow so fast, but in the end achieves higher levels of accuracy than l1 .
Which of the two curves indicates a preferable learner depends on the circum-
stances of the given application. When the source of the training examples is limited,
the first learner is clearly more appropriate. If the examples are abundant, we will
prefer the other learner, assuming of course that the computational costs are not
prohibitive.
Computational Costs There are two aspects to computational costs. First is the
time needed for the induction of the classifier from the available data. The second is
the time it takes to classify a set of examples with the classifier thus induced.
Along these lines, the techniques described in this book cover a fairly broad
spectrum. As for induction costs, the cheapest is the basic version of the k-NN
classifier: the only “computation” involved is to store the training examples.2 On
the other hand, the k-NN classifier is expensive in terms of the classification costs.
For instance, if we have a million training examples, each described by ten thousand
attributes, then tens of billions of arithmetic operations will have to be carried out to
classify a single example. When asked to classify millions of examples, even a very
fast computer will take quite some time.
The situation is different in the case of decision trees. These are cheap when
used to classify examples: usually only a moderate number of single-attribute tests
are needed. However, induction of decision trees can take a lot of time if many
training examples described by many attributes are available.
Induction costs and classification costs of the other classifiers vary, and the
engineer is well advised to consider these costs when choosing the most appropriate
machine-learning paradigm for the given application. Just as important is a solid
understanding of how these costs depend on the number of the training examples,
on the number of attributes describing them, and sometimes also on the required
accuracy (e.g., in the case of neural networks).
• What are the two main aspects of costs associated with a given machine-learning
paradigm?
• What does the learning curve tell us about a machine learning algorithm’s
behavior? What shape of the learning curve do we want in the ideal case? What
should we expect in reality?
• Under what circumstances is a steeper learning curve with lower maximum better
than a flatter curve with a higher maximum?
2
Some computation will be necessary if we decide to remove noisy or redundant examples.
11.5 Methodologies of Experimental Evaluation
The reader understands that different domains will benefit from different induction
techniques. The choice is usually not difficult to make. Some knowledge of the
available training data often helps us choose the most appropriate paradigm; for
instance, if a high percentage of the attributes are suspected to be irrelevant, then
decision trees are likely to be more successful than a nearest-neighbor classifier.
However, the success of a given technique also depends on the values of various
parameters. Although even here certain time-tested rules of thumb can help, the best
parameter setting is usually found only by experimentation.
Baseline Approach and Its Limitations The basic scenario is simple. The set of
pre-classified examples is divided into two subsets, one used for training, the other
for testing. The training-testing session is repeated for different parameter settings,
and the one that results in the highest performance is then chosen.
This, however, can only be done if a great many pre-classified examples are
available. In domains where examples are scarce, or expensive to obtain, a random
division into a pair of the training and testing sets will lack objectivity. Either
of the two sets can, by mere chance, misrepresent the given domain adequately.
Statisticians are telling us that both the training set and the testing set should have
more or less the same distribution of examples. In small sets, of course, this cannot
be guaranteed.
Random Subsampling When the set of pre-classified examples is small, the
engineer usually prefers to repeat the training-testing procedure several times. In
each run, the set of examples is randomly divided into two parts, one for training,
the other for testing. The measured performances are recorded, and then averaged.
Care has to be taken that the individual data-splits are mutually independent.
Once the procedure has been repeated N times (typically, N D 10 or N D 5), the
results are reported in terms of the average classification accuracy and the standard
deviation, say, 84:2 ˙ 0:6. For the calculation of averages and standard deviations,
the formulas from Chap. 2 are used: the average, , is obtained by Eq. (2.12);
the standard deviation, , is obtained as the square root of variance, 2 , which is
calculated using Eq. (2.13).
N-Fold Crossvalidation For more advanced statistical evaluation, experienced
experimenters often prefer the so-called N-fold crossvalidation. The principle is
shown in Fig. 11.4. To begin with, the set of pre-classified examples is divided into
N equally sized (or almost equally-sized) subsets which the machine-learning jargon
sometimes (not quite correctly) refers to as “folds.”
N-fold crossvalidation then runs N experiments. In each, one of the N subsets
is removed so as to be used only for testing (this guarantees that, in each run, a
different testing set is used). The training is then carried out on the union of the
remaining N 1 subsets. Again, the results are averaged, and the standard deviation
calculated.
11.5 Methodologies of Experimental Evaluation 225
Fig. 11.4 N-fold 5−fold cross−validation

cross-validation divides the
training set into N equally 1.
sized subsets. In each of the
N experimental runs, it 2. ... testing
withholds a different subset
for testing, inducing the 3.
classifier from the union of 4. ... training
the remaining N 1 subsets
5.
the original trainig set
The advantage of N-fold crossvalidation as compared to random subsampling is

that the testing sets are disjoint (non-overlapping), which is deemed advantageous
for certain types of statistical evaluations (see Chap. 12).
Stratified Approaches Suppose you are dealing with a domain with 60 positive
and 940 negative examples. If you rely on N-fold crossvalidation with N D 10,
each of the “folds” is likely to contain a different number of positive examples. On
average, there will be six positives in each fold, but the concrete numbers can vary
significantly. In a domain where one of the classes is relatively rare, it can even
happen that some of the “folds” will not contain any single positive example at all.
The situation is typically encountered in domains with heavily imbalanced class
representation.
In this event, the experienced experimenter prefers a so-called stratified approach
to N-fold crossvalidation. The idea is to make sure that each of the N “folds” con-
tains (approximately) the same representation of the examples from the individual
classes. For instance, when using the fivefold crossvalidation in a domain with 60
and 940 positive and negative examples, respectively, each fold should consist of
200 examples of which 12 are positive.
The same principle is often used in the random-subsampling approach—which,
admittedly, is in its stratified version no longer totally random. Again, the idea
is to make sure that each training set, and each testing set, has about the same
representation of each class.
5 2 Crossvalidation (5 2cv) There is yet another approach to experimental
evaluation of machine-learning techniques, the so-called 5 2 crossvalidation,
sometimes abbreviated as 5 2cv. This may actually be the most popular method-
ology in machine learning. The principle is built around a combination of random
subsampling and twofold crossvalidation.
To be more specific, 5 2cv divides the set of pre-classified examples into two
equally sized parts, T1 and T2 . Next, it uses T1 for training and T2 for testing, and
then the other way round: T2 for training and T1 for testing. The procedure is then
repeated five times, each time with a different random division into two subsets. All
Table 11.3 The algorithm for 5 2 cross-validation (5 2 CV)

Let T be the original set of pre-classified examples.
1. Divide T randomly into two equally-sized subsets. Repeat the division five times. The result
is five pairs of subsets denoted as Ti1 and Ti2 (for i D 1; : : : 5).
2. For each of these pairs, use Ti1 for training and Ti2 for testing, and then the other way
round.
3. For the ten training/testing sessions thus obtained, calculate the mean value and the
standard deviation of the chosen performance criterion.
in all, ten learning/testing sessions are thus created. The principle is summarized by
the pseudocode in Table 11.3.
Again, many experimenters prefer to work with the stratified version of this
methodology, making sure that the representation of the individual classes is about
the same in each of the ten parts used in the experiments.
The No-Free-Lunch Theorem It would be foolish to expect some machine-
learning technique to be a holy grail, a mechanism to be preferred under all
circumstances. Nothing like this exists. The reader by now understands that
each paradigm has its advantages that make it succeed in some domains—and
shortcomings that make it fail miserably in others. Only systematic experiments can
tell the engineer which type of classifier, and which induction algorithm, to select
for the task at hand. The truth of the matter is that no machine-learning approach
will outperform all other machine-learning approaches under all circumstances.
Mathematicians have been able to prove the validity of this statement by a
rigorous proof. The result is known under the (somewhat fancy) name of “no-free-
lunch theorem.”
• What is the difference between N-fold cross-validation and random subsam-
pling? Why do we sometimes prefer to employ the stratified versions of these
methodologies?
• Explain the principle of the 52 cross-validation (52 cv), including its stratified
version.
• What does the so-called no-free-lunch theorem tell us?
• The basic criterion to measure classification performance is error rate, E, defined

as the percentage of misclassified examples in the given set. The complementary
quantity is classification accuracy, Acc D 1 E.
• When the evidence for any class is not sufficiently strong, the classifier should
better reject the example to avoid the danger of a costly misclassification.
Rejection rate then becomes yet another important criterion for the evaluation
of classification performance. Higher rejection rate usually means lower error
rate; beyond a certain point, however, the classifier’s utility will degrade.
• Criteria for measuring classification performance can be defined by the counts
(denoted as NTP ; NTN ; NFP ; NFN ;) of true positives, true negatives, false positives,
and false negatives, respectively.
• In domains with imbalanced class representation, error rate can be a misleading
criterion. A better picture is offered by the use of precision (Pr D NTPNCN TP
FP
) and
recall (Re D NTP CNFN ).
NTP
• Sometimes, precision and recall are combined in a single criterion, Fˇ , that is

defined by the following formula:
.ˇ 2 C 1/ Pr Re
Fˇ D
ˇ 2 Pr C Re
The value of the user-set parameter ˇ determines the relative importance of

precision (ˇ < 1) or recall (ˇ > 1). When the two are deemed equally important,
we use ˇ D 1, obtaining the following:
2 Pr Re
F1 D
Pr C Re
• Less common criteria for classification performance include sensitivity, speci-

ficity, and gmean.
• In domains where an example can belong to more than one class at the same time,
the performance is often evaluated by an average taken over the performances
measured along the individual classes. Two alternative methods of averaging are
used: micro-averaging and macro-averaging.
• Another important aspect of a machine-learning technique is how many training
examples are needed if a certain classification performance is to be reached. The
situation is sometimes visualized by means of a learning curve. Also worth the
engineer’s attention are the computational costs associated with induction and
with classification.
• When comparing alternative machine-learning techniques in domains with lim-
ited numbers of pre-classified examples, engineers rely on methodologies known
as random subsampling, N-fold cross-validation, and the 5 2 cross-validation.
The stratified versions of these techniques make sure that each training set (and
testing set) has the same proportion of examples for each class.
Historical Remarks Most of the performance criteria discussed in this chapter are
well established in the statistical literature, and have been used for such a long time
that it is difficult to trace their origin. The exception is the relatively recent gmean
that was proposed to this end by Kubat et al. [51].
The idea to refuse to classify examples where the k-NN classifier cannot rely
on a significant majority was put forward by Hellman [36] and later analyzed by
Louizou and Maybank [56]. The principle of 5 2 cross-validation was suggested,
and experimentally explored, by Dietterich [22]. The no-free-lunch theorem was
published by Wolpert [100].
Exercises
1. Suppose that the evaluation of a classifier on a testing set resulted in the counts
summarized in the following table:
Labels
pos neg
neg 40 850
Calculate the values of precision, recall, sensitivity, specificity, and gmean.

2. Using the data from the previous question, calculate Fˇ for different values of
the parameter: ˇ D 0:5; ˇ D 1, and ˇ D 2.
3. Suppose that an evaluation of a machine-learning technique using fivefold cross-
validation resulted in the following error rates measured in the testing sets:
E11 D 0:14; E12 D 0:16; E13 D 0:10; E14 D 0:15; E15 D 0:18
E21 D 0:17; E22 D 0:15; E23 D 0:12; E24 D 0:13; E25 D 0:20
Calculate the mean value of the error rate as well as the standard deviation, ,
using the formulas from Chap. 2 (do not forget that standard deviation is the
square root of variance, 2 ).
1. Suggest a domain where precision is much more important than recall; con-
versely, suggest a domain where it is the other way round, recall being more
important than precision.
(Of course, use different examples than those mentioned in this chapter.)
2. What aspects of the given domain is reflected in the pair, sensitivity and
specificity? Suggest circumstances under which these two give a better picture
of the classifier’s performance than precision and recall.
3. Suppose that, for a given domain, you have induced two classifiers: one with
very high precision, the other with high recall. What can be gained from the
combination of the two classifiers? How would you implement this combination?
Under what circumstances will the idea fail?
4. Try to think about the potential advantages and shortcomings of random subsam-
pling in comparison with N-fold crossvalidation.
1. Assume that some machine-learning experiment resulted in a table where each

row represents a testing example. The first column contains the examples’ class
labels (“1” or “0” for the positive and negative examples, respectively), and the
second column contains the labels suggested by the induced classifier.
Write a program that calculates precision, recall, as well as Fˇ for a user-
specified ˇ.
Write a program that calculates the values of the other performance criteria.
2. Suppose that the training set has the form of a matrix where each row represents
an example, each column represents an attribute, and the rightmost column
contains the class labels.
Write a program that divides this set randomly into five pairs of equally sized
subsets, as required by the 5 2 cross-validation technique. Then write another
program that creates the subsets in the stratified manner where each subset has
approximately the same representation of each class.
3. Write a program that accepts two inputs: (1) a set of class labels of multi-label
testing examples, and (2) the labels assigned to these examples by a multi-label
classifier. The output consists of micro-averaged and macro-averaged values of
precision and recall.
4. Write a computer program that accepts as input a training set, and outputs N
subsets to be used in N-fold crossvalidation. Make sure the approach is stratified.
How will your program have to be modified if you later decide to use the 5 2
crossvalidation instead of the plain N-fold crossvalidation?
Chapter 12
Statistical Significance
Suppose you have evaluated a classifier’s performance on an independent testing set.

To what extent can you trust your findings? When a flipped coin comes up heads
eight times out of ten, any reasonable experimenter will suspect this to be nothing
but a fluke, expecting that another set of ten tosses will give a result closer to reality.
Similar caution is in place when measuring classification performance. To evaluate
classification accuracy on a testing set is not enough; just as important is to develop
some notion of the chances that the measured value is a reliable estimate of the
classifier’s true behavior.
This is the kind of information that an informed application of mathematical
statistics can provide. To acquaint the student with the requisite techniques and
procedures, this chapter introduces such fundamental concepts as standard error,
confidence intervals, and hypothesis testing, explaining and discussing them from
the perspective of the machine-learning task at hand.
12.1 Sampling a Population
If we test a classifier on several different testing sets, the error rate on each of them
will be different—but not totally arbitrary: the distribution of the measured values
cannot escape the laws of statistics. A good understanding of these laws can help us
estimate how representative the results of our measurements really are.
An Observation Table 12.1 contains one hundred zeros and ones, generated by
a random-number generator whose parameters have been set to make it return
a zero 20% of a time, and a one 80% of the time. The real percentages in the
generator’s output are of course slightly different than what the setting required.
In this particular case, the table contains 82 ones and 18 zeros.

DOI 10.1007/978-3-319-63913-0_12
232 12 Statistical Significance
Table 12.1 A set of binary 0 0 1 0 1 1 1 0 1 1 6

values returned by a
1 1 0 1 1 1 1 1 1 1 9
random-number generator set
to return a one 80% of the 1 1 1 0 1 1 1 1 1 1 9
time 1 1 1 1 1 1 0 0 1 1 8
1 1 1 0 1 0 1 0 1 1 7
1 1 1 1 1 1 1 1 1 1 10
1 1 1 1 1 1 1 1 0 1 9
1 1 1 0 1 1 1 0 1 1 8
1 1 1 0 1 0 1 1 1 1 8
1 0 1 1 1 1 1 0 1 1 8
9 8 9 5 10 8 9 5 9 10 82
In reality, there are 82 ones and 18 zeros. At
the ends of the rows and columns are the cor-
responding sums
The numbers on the side and at the bottom of the table tell us how many ones are
found in each row and column. Based on these, we can say that the proportions of
ones in the first two rows are 0.6 and 0.9, respectively, because each row contains
10 numbers. Likewise, the proportions of ones in the first two columns are 0.9 and
0.8. The average of these four proportions is .0:6 C 0:9 C 0:9 C 0:8/=4 D 0:80, and
the standard deviation is 0:08.1
For a statistician, each row or column represents a sample of the population. All
samples have the same size: n D 10. Now, suppose we increase this value to, say,
n D 30. How will the proportions be distributed then?
Returning to the table, we can see that the first three rows combined contain
6 C 9 C 9 D 24 ones, the next three rows contain 8 C 7 C 10 D 25 of them,
the first three columns contain 9 C 8 C 9 D 26, and the next three columns contain
5C10C8 D 23. Dividing each of these numbers by n D 30, we obtain the following
proportions: 2430
D 0:80; 25
30
D 0:83; 26
30
D 0:87, and 23 30
D 0:77. Calculating the
average and the standard deviation of these four values, we get 0:82 ˙ 0:02.
If we compare the results observed in the case of n D 10 with those for
n D 30, we notice two things. First, there is a minor difference between the
average calculated for the bigger samples (0.82) versus the average calculated for
the smaller samples (0.80). Second, the bigger samples exhibit a clearly smaller
standard deviation: 0.02 for n D 30 versus 0.08 for n D 10. Are these observations
explained by mere coincidence, or are they the consequence of some underlying
law?
1
Recall that standard deviation is the square root of variation; this, in turn, is calculated by
Eq. (2.13) from Chap. 2.
12.1 Sampling a Population 233
Estimates Based on Random Samples The answer is provided by a theorem

that says that estimates based on samples become more accurate with the growing
sample size, n. Further on, the larger the samples, the smaller the variation of the
estimates from one sample to another.
Another theorem, the so-called central limit theorem, states that the distribution
of the individual estimates can be approximated by the Gaussian normal distribution
which we already know from Chap. 2—the reader will recall its signature bell-like
shape. However, this approximation is known to be reasonably accurate only if the
proportion, p, and the sample size, n, satisfy the following two conditions:
np 10 (12.1)
n.1 p/ 10 (12.2)
If the conditions are not satisfied (if at least one of the products is less than 10),
the distribution of estimates obtained from the samples cannot be approximated by
the normal distribution without certain loss in accuracy.
Sections 12.2 and 12.3 will elaborate on how the normal-distribution approxi-
mation can help us establish our confidence in the measured performance of the
induced classifiers.
An Illustration Let us check how these conditions are satisfied in the case of
the samples of Table 12.1. We know that the proportion of ones in the original
population was determined by the user-set parameter of the random-number
generator: p D 0:8. Let us begin with samples of size n D 10. It turns out that
none of the two conditions is satisfied because np D 10 0:8 D 8 < 10 and
n.1 p/ D 10 0:2 D 2 < 10. Therefore, the distribution of the proportions
observed in these small samples cannot be approximated by the normal distribution.
In the second attempt, the sample size was increased to n D 30. As a result, we
obtain np D 30 0:8 D 24 > 10, and this means that Condition 12.1 is satisfied.
At the same time, however, Condition (12.2) is not satisfied because n.1 p/ D
30 0:2 D 6 < 10. Even here, therefore, the normal distribution does not offer
sufficiently accurate approximation.
The situation will change if we increase the sample size to n D 60. Doing the
math, we easily establish that np D 600:8 D 48 10 and also n.1p/ D 600:2 D
12 10. We can therefore conclude that the distribution of the proportions of ones
in samples of size n D 60 can be approximated with the normal distribution without
any perceptible loss in accuracy.
The Impact of p Note how the applicability of the normal distribution is affected
by p, the proportion of ones in the entire population. It is easy to see that, for
different values of p, different sample sizes are called for if the two conditions are to
be satisfied. Relatively small size is sufficient if p D 0:5; but the more the proportion
differs from p D 0:5 to either side, the bigger the samples that we need.
To get a better idea of what this means in practice, recall that we found the sample
size of n D 60 to be sufficient in a situation where p D 0:8. What if, however, we
decide to base our estimates on samples of the same size, n D 60, but in a domain
where the proportion is higher, say, p D 0:95? In this event, we will realize that
n.1 p/ D 60 0:05 D 3 < 10, which means that Condition (12.2) is not met, and
the distribution of the proportions in samples of this size cannot be approximated
by the normal distribution. For this condition to be satisfied in this domain, we
would need a sample size of at least n D 200. Since 200 0:05 D 10, we have
just barely made it. By the way, note that, on account of the symmetry of the two
conditions, (12.1) and (12.2), the same minimum size, n D 200, will be called for
in a domain where p D 0:05 instead of p D 0:95.
Parameters of the Distribution Let us return to our attempt to estimate the
proportion of ones based on sampling. We now know that if the samples are large
enough, the distribution of estimates made in different samples can be approximated
by the normal distribution whose mean equals the (theoretical) proportion of ones
that would have been observed in the entire population if such an experiment were
possible.
The other parameter of a distribution is the standard deviation. In our context,
statisticians prefer the term standard error, a terminological subtlety essentially
meant to indicate the following: whereas “standard deviation” refers to a distribution
of any variable (such as weight, age, or temperature), the term “standard
error” is used when we refer to variations of estimates from one sample to another.
And this is what interests us in the case of our proportions.
Let us denote the standard error by sE . Mathematicians have established that its
value can be calculated from the sample size, n, and the theoretical proportion, p,
using the following formula:
r
p.1 p/
sE D (12.3)
n
For instance, if n D 50 and p D 0:80, then the standard error is as follows:

r
0:80 0:20
sE D D 0:06
50
When expressing this result in plain English, some engineers prefer to say that
the standard error is 6%.
The Impact of n; Diminishing Returns Note how the value of the standard error
goes the other way than the sample size, n. To be more specific, the larger the
samples, the lower the standard error and vice versa. Thus in the case of n D 50
and p D 0:80, we obtained sE Dq 0:06. If we use larger samples, say, n D 100, the
standard error will drop to sE D 0:80:2
100
D 0:04. The curve defined by the normal
distribution thus becomes narrower, and the proportions in different samples will
tend to be closer to p.
12.2 Benefiting from the Normal Distribution 235
This said, we should also be aware of the fact that increasing the sample size
brings diminishing returns. Let us illustrate this statement using a simple example.
The calculations carried out in the previous paragraph convinced us that, when
proceeding from n D 50 to n D 100 (doubling the sample size), we managed to
reduce sE by two percentage points, from 6 to 4%. If, however, we do the same
calculation for n D 1000, we get sE D 0:013, whereas n D 2000 results in
sE D 0:009. In other words, doubling the sample size from 1000 to 2000, we only
succeeded in reducing the standard error from 1.3 to 0.9%, which means that the
only reward for doubling the sample size was the paltry 0.4%.
This last observation is worth remembering—for very practical reasons. In many
domains, pre-classified examples are difficult or expensive to obtain; the reader
will recall that this was the case of the oil-spill domain discussed in Sect. 8.2. If
acceptable estimates of proportions can be made using a relatively small testing
set, the engineer will not want to go into the trouble of trying to procure additional
examples; the miniscule benefits may not justify the extra costs.
• Write down the formulas defining the conditions to be satisfied if the distribution
of the proportions obtained from random samples are to follow the normal
distribution.
• Explain how the entire-population proportion, p, affects the sample size, n, that is
necessary for the proportions measured on different samples to follow the normal
distribution.
• What is the mean value of a set of estimates that have been made based on
different samples? Also, write down the formula that calculates the standard
error.
• Elaborate on the statement that “increasing the sample size brings only diminish-
ing returns.”
12.2 Benefiting from the Normal Distribution
The previous section investigated the proportions of ones in samples taken from a
certain population. The sample size was denoted by n, and the theoretical proportion
of ones in the whole population was denoted by p. This theoretical value we do not
know; the best we can do is estimate it based on our observation of a sample. Also,
we have learned that, while the proportion in each individual sample is different, the
Fig. 12.1 Gaussian (normal)

distribution whose mean
value is p
−3σ −2σ −σ p σ 2σ 3σ
distribution of these values can often be approximated by the normal distribution—

the approximation being reasonably accurate if Conditions (12.1) and (12.2) are
satisfied.
The normal distribution can help us decide how much to trust the classification
accuracy (or, for that matter, any other performance criterion) that has been
measured on one concrete testing set. To be able to do so, let us take a brief look at
how to calculate so-called confidence values.
Re-formulation in Terms of a Classifier’s Performance Suppose the ones and
zeros in Table 12.1 represent correct and incorrect classifications, respectively, as
they have been made by a classifier being evaluated on a testing set that consists of
one hundred examples (one hundred being the number of entries in the table). In this
event, the proportion of ones gives the classifier’s accuracy, whereas the proportion
of zeros defines its error rate.
Evaluation of the classifier on a different testing set will of course result in
different values of the classification accuracy or error rate. But when measured on
great many testing sets, the individual accuracies will be distributed in a manner
that, as we have seen, roughly follows the normal distribution.
Properties of the Normal Distribution Figure 12.1 shows the fundamental shape
of the normal distribution. The vertical axis represents the probability density
function as we know it from Chap. 2. The horizontal axis represents classification
accuracy. The mean value, denoted here as p, is the theoretical classification
accuracy which we would obtain if we had a chance to evaluate the classifier
on all possible examples from the given domain. This theoretical value is of course
unknown, which is why our intention is to estimate it on the basis of a concrete
sample—the available set of testing examples.
The bell-like shape of the density function reminds us that most testing sets
will yield classification accuracies relatively close to the mean, p. The greater
the distance from p, the smaller the chance that this particular performance will
be obtained from a random testing set. Note also that, along the horizontal axis,
the graph highlights certain specific distances from p: the multiples of , the
distribution’s standard deviation—or, when we deal with sample-based estimates,
the standard error of these estimates.
12.2 Benefiting from the Normal Distribution 237
Table 12.2 For the normal Confidence level (%) z

distribution, with mean p and
standard deviation , the left 68 1.00
column gives the percentage 90 1.65
of values found in the interval 95 1.96
Œ p z ; p C z 98 2.33
99 2.58
The formula defining the normal distribution was introduced in Sect. 2.5 where it
was called the Gaussian “bell” function. Knowing the formula, we can establish
the percentage of values found within a specific interval, Œa; b. The size of the
entire area under the curve (from minus infinity to plus infinity) is 1. Therefore,
if the area under the curve within the range of Œa; b is 0.80, we can say that 80% of
the performance estimates are found in this interval.
Identifying Intervals of Interest Not all intervals are equally important. For the
needs of classifier evaluation, we are interested in those that are centered at the
mean value, p. For instance, the engineer may want to know what percentage of
values will be found in Œ p ; p C . Conversely, she may want to know the size
of the interval (again, centered at p) that contains 95% of all values.
Strictly speaking, questions of this kind can be answered with the help of
mathematical analysis. Fortunately, we do not need to do the math ourselves because
others have done it before, and we can take advantage of their findings. Some of the
most useful results are shown in Table 12.2. Here, the left column lists percentages
called confidence levels; for each of these, the right column specifies the interval
that comprises the given percentage of values. Note that the length of the interval
is characterized by z , the number of standard deviations to either side of p. More
formally, therefore, the interval is defined as Œ p z ; p C z .
Here is how the table is used for practical purposes. Suppose we want to know
the size of the interval that contains 95% of the values. This percentage is found in
the third row. We can see that the number on the right is 1.96, and this is interpreted
as telling us that 95% of the values are in the interval Œ p 1:96 ; p C 1:96 .
Similarly, 68% of the values are found in the interval Œ p ; p C —this is what
we learn from the first row in the table.
Standard Error of Sample-Based Estimates Let us take a look at how to employ
this knowledge when evaluating classification accuracies. Suppose that the testing
sets are all of the same size, n, and suppose that this size satisfies Conditions (12.1)
and (12.1) that allow us to use the normal distribution. We already know that the
average of the classification accuracies measured on great many independent testing
sets will converge to the theoretical accuracy, the one that would have been obtained
by testing the classifier on all possible examples.
The standard error2 is calculated using Eq. (12.3). For instance, if the theoretical
classification accuracy is p D 0:70, and the size of each testing set is n D 100, then
the standard error of the classification accuracies obtained from great many different
testing sets is calculated as follows:
r r
p.1 p/ 0:7.1 0:7/
sacc D D D 0:046 (12.4)
n 100
After due rounding, we will say that the classification accuracy is 70% plus or
minus 5%. Note, again, that the standard error will be lower if we use a larger testing
set. This makes sense: the larger the testing set, the more thorough the evaluation,
and thus the higher our confidence in the value thus obtained.
Let us now ask what value we are going to obtain if we evaluate the classifier
on some other testing sets of the same size. Once again, we answer the question
with the help of Table 12.2. First of all, we find the row representing 95%. In this
row, the right column gives the value z D 1:96; and this is interpreted as telling
us that 95% of all results will be in the interval Œ p 1:96 sacc ; p C 1:96 sacc D
Œ0:80 1:96 0:46; 0:80 C 1:96 0:46 D Œ0:61; 0:79.
Do not forget, however, that this will only be the case if the testing set has the
same size, n D 100. For a different n, Eq. (12.4) will give us a different standard
error, sacc , and thus a different interval.
Two Important Reminders It may be an idea to remind ourselves of what exactly
the normal-distribution assumption is good for. Specifically, if the distribution is
normal, then we can use Table 12.2 from which we learn the size of the interval
(centered a p) that contains the given percentage of values.
On the other hand, the formula for standard error (Eq. (12.3)) is valid generally,
even if the distribution is not normal. For the calculation of standard error, the two
conditions, (12.1) and (12.2), do not have to be satisfied.
• How do the considerations from the previous section apply to the evaluation of
an induced classifier’s performance?
• What kind of information can we glean from Table 12.2? How can this table
be used when quantifying the confidence in the classification-accuracy value
obtained from a testing set of size n?
2
As explained in Sect. 12.1 in connection with the distribution of results obtained from different
samples, we prefer the term standard error to the more general standard deviation.
12.3 Confidence Intervals 239
• How will you calculate the standard error of estimates based on a given testing
set? How does this standard error depend on the size of the testing set?
12.3 Confidence Intervals
Let us now focus on how the knowledge gained in the previous two sections can help
us specify the experimenter’s confidence in the classifier’s performance as measured
on the given testing data.
Confidence Interval: An Example Now that we understand how the classification
accuracies obtained from different testing sets are distributed, we are ready to
draw conclusions about how confident we can be in our expectation that the value
measured on one concrete testing set is close to the true theoretical value.
Suppose the size of the testing set is n D 100, and let the classification accuracy
measured on this testing set be acc D 0:85. For the training set of this size, the
standard error is as follows:
r
0:85 0:15
sacc D D 0:036 (12.5)
100
Checking the normal-distribution conditions, we realize that they are both

satisfied here because 100 0:85 D 85 10 and 100 0:15 D 15 10. This
means that we can take advantage of the z -values listed in Table 12.2. Using
this table, we easily establish that 95% of all values are found in the interval
Œacc 1:96 sacc ; acc C 1:96 sacc . For acc D 0:85 and sacc D 0:036, we realize
that the corresponding interval is Œ0:85 0:07; 0:85 C 0:07 D Œ0:78; 0:92.
What this result is telling us is that, based on the evaluation on the given testing
set, we can say that, with 95% confidence, the real classification accuracy finds itself
somewhere in the interval Œ0:78; 0:92. This interval is usually called the confidence
interval.
Two New Terms: Confidence Level and Margin of Error Confidence intervals
reflect specific confidence levels—those defined by the percentages listed in the left
column of Table 12.2. In our specific case, the confidence level was 95%.
Each confidence level defines a different confidence interval. This interval can
be re-written as p ˙ M where p is the mean and M is the so-called margin of error.
For instance, in the case of the interval Œ0:78; 0:92, the mean was p D 0:85 and the
margin of error was M D z sacc D 1:96 0:036 D 0:07.
Choosing the Confidence Level In the example discussed above, the requested
confidence level was 95%, a fairly common choice. For another confidence level, a
different confidence interval would have been obtained. Thus for 99%, Table 12.2
gives z D 2:8, and the confidence interval is Œ0:852:58sacc ; 0:85C2:58sacc D
Œ0:76; 0:94. Note that this interval is longer than the one for confidence level 95%.
This was to be expected: the chance that the real, theoretical, classification accuracy
finds itself in a longer interval is higher. Conversely, it is less likely that the
theoretical value will fall into some narrower interval. Thus for the confidence level
of 68% (and the standard error rounded to sacc D 0:04), the confidence interval is
Œ0:85 0:04; 0:85 C 0:04 D Œ0:81I 0:89.
Importantly, we must not forget that, even in the case of confidence level
99%, one cannot be absolutely sure that the theoretical value will fall into the
corresponding interval. There is still that 1% probability that the measured value
will be outside this interval.
Another Parameter: Sample Size The reader now understands that the length of
the confidence interval depends on the standard error, and that the standard error, in
turn, depends on the size, n, of the testing set (see Eq. (12.3)). Essentially, the larger
the testing set, the stronger the evidence in favor of the measured value, and thus the
narrower the confidence interval. This is why we say that the margin of error and the
training-set size are in inverse relation: as the training-set size increases, the margin
of error decreases.
Previously, we mentioned that a higher confidence level results in a longer
confidence interval. If we think this interval to be too big, we can make it shorter by
using a bigger testing set, and thus a higher value of n (which decreases the value of
the standard error of the measured value).
There is a way of deciding how large the testing set should be if we want to limit
the margin of error to a certain maximum value. Here is the formula calculating the
margin of error:
r
p.1 p/
M D z sacc D z (12.6)
n
Solving this equation for n (for specific values of M. p, and z ) will give us the
required testing-set size.
A Concluding Remark The method of establishing the confidence interval for
the given confidence level was explained using the simplest performance criterion,
classification accuracy. Yet the scope of the method’s applicability is much broader:
the uncertainty of any variable that represents a proportion can thus be quantified.
In the context of machine learning, we can use the same approach to establish our
confidence in any of the performance criteria from Chap. 11, be it precision, recall,
or some other quantity.
But we have to be careful we do it right. For one thing, we must not forget that
the distribution of the values of the given quantity can only be approximated by the
normal distribution if the conditions (12.1) and (12.2) are satisfied. Second, we have
to make sure we understand the meaning of n when calculating the standard error
using Eq. (12.3). For instance, the reader remembers that precision is calculated with
the formula, NTPNCNTP
FP
: the percentage of true positives among all examples labeled
by the classifier as positive. This means that we are dealing with a proportion of true
positives in a sample of the size n D NTP C NFP . Similar considerations have to be
made in the case of recall and some other performance criteria.
12.4 Statistical Evaluation of a Classifier 241
• Explain the meaning of the term, confidence interval. What is meant by the
margin of error?
• How does the size of the confidence interval (and the margin of error) depend
on the user-specified confidence level? How does it depend on the size of the
testing set?
• Discuss the calculations of confidence intervals for some other performance
criteria such as precision and recall.
12.4 Statistical Evaluation of a Classifier
A claim about a classifier’s performance can be confirmed or refuted experimentally,

by testing the classifier on a set of pre-classified examples. One possibility for the
statistical evaluation of the results thus obtained is to follow the algorithm from
Table 12.3. Let us illustrate the procedure on a simple example.
A Simple Example Suppose a machine-learning specialist tells you that the
classifier he has induced has classification accuracy acc D 0:78. Faithful to the
dictum, “trust but verify,” you decide to find out whether this statement is correct.
To this end, you prepare n D 100 examples whose class labels are known, and then
set about measuring the classifier’s performance on this testing set.
Table 12.3 The algorithm for statistical evaluation of a classifier’s performance
1. For the given size, n, of the testing set, and for the claimed classification accuracy, acc,
check whether the conditions for normal distribution are satisfied:
n acc 10 and n .1 acc/ 10
2. Calculate the standard error by the usual formula:

r
acc.1 acc/
sacc D
n
3. Assuming that the normal-distribution assumption is correct, find in Table 12.2 the z -value
for the requested level of confidence. The corresponding confidence interval is Œacc z
sacc ; acc C z sacc .
4. If the value measured on the testing set finds itself outside this interval, reject the claim that
the accuracy equals acc. Otherwise, assume that the available evidence is insufficient for
the rejection.
Let the experiment result in giving us classification accuracy 0:75. Well, this
is less than the promised 0:78, but then: is this observed difference still within
reasonable bounds? To put it another way, is there a chance that the specialist’s
claim was correct, and that the lower performance measured on the testing set can
be explained by the variations implied by the random nature of the employed testing
data? After all, a different testing set is likely to result in a different classification
accuracy.
Checking the Conditions for Normal Distribution The first question to ask is
whether the distribution of the performances thus obtained can be approximated
by the normal distribution. A positive answer will allow us to base our statistical
evaluation on the values from Table 12.2.
Verification of Conditions (12.1) and (12.2) is quite easy. Seeing that np D 100
0:75 D 75 10, and that n.1 p/ D 100 0:25/ D 25 10, we realize that the
conditions are satisfied and the normal-distribution assumption can be used.
Finding the Confidence Interval for the 95%-Confidence Level Suppose that
you are prepared to accept the specialist’s claim (acc D 0:78) if there is at least
a 95% chance that such performance will make it possible to observe that the
classification accuracy on a random testing set is acc D 0:75. This will be possible
if 0.75 finds itself within the corresponding confidence interval, centered at 0.78.
Let us find out whether this is the case.
The corresponding row in the table informs us that z D 1:96; this means that
95% of accuracies obtained on random testing set will find themselves in the interval
Œacc 1:96 sacc ; acc C 1:96 sacc , where acc D 0:78 is the original claim and sacc
is the standard error to be statistically expected for testing sets of the given size, n.
In our concrete training set, the size is n D 100. The standard error is calculated
as follows:
r r
acc.1 acc/ 0:75 0:25
sacc D D D 0:043 (12.7)
n 100
We conclude that the confidence interval is Œ0:781:960:043; 0:78C1:960:043

which, after evaluation and due rounding, is Œ0:70; 0:86.
A Conclusion Regarding the Specialist’s Claim Evaluation on our own training
set resulted in classification accuracy acc D 0:75, a value that finds itself within the
confidence interval corresponding to the chosen confidence level of 95%.
This is encouraging. For the given claim, acc D 0:78, there is a 95% probability
that our evaluation on a random testing set will give us a classification accuracy
somewhere within the interval Œ0:70; 0:86. This, indeed, is what happened in this
particular case. And so, although our result, acc D 0:75, is somewhat lower than
the specialist’s claim, we have to admit that our experimental evaluation failed to
provide convincing evidence against the claim. In the absence of such evidence, we
accept the claim as valid.
12.4 Statistical Evaluation of a Classifier 243
Type-I Error in Statistical Evaluation: False Alarm The reader now understands
the fundamental principle of statistical evaluation. Someone makes a statement
about performance. Based on the size of our testing set (and assuming normal
distribution), we calculate the size of the interval that is supposed to contain the
given 95% of all values. There is only a 5% chance that, if the original claim is
correct, the result of testing will be outside this interval. This is why we reject
any hypothesis whose testing results landed in this less-than-5% region. We simply
assume that it is rather unlikely that such difference would be observed.
This said, such difference should still be expected in 5% cases. We have to admit
that there exists some small danger that the evaluation of the classifier on a random
testing set will result in a value outside the given confidence interval. In this case,
rejecting the specialist’s claim would be unfair. Statisticians call this the type-I error:
the false rejection of an otherwise correct claim; a rejection that is based on the fact
that certain results are untypical.
If we do not like to face this danger, we can reduce it by increasing the required
confidence level. If we choose 99% instead of the 95%, false alarms will be less
frequent. But this reduction is not gained for free—as will be explained in the next
paragraph.
Type-II Error in Statistical Evaluation: Failing to Detect an Incorrect Claim
Also the opposite case is possible. To wit, the initial claim is false, and yet the
classification accuracy obtained from our testing falls within the given confidence
interval. When this happens, we are forced to conclude that our experiment failed to
provide sufficient evidence against the claim; the claim thus has to be accepted.
The reader may find this unfortunate, but this is indeed what sometimes happens.
An incorrect claim is not refuted. Statisticians call this the type-II error. It is
typical of those cases where very high confidence level is required: so broad is the
corresponding interval that the results of testing will almost never fall outside; the
experimental results then hardly ever lead to the rejection of the initial claim.
The thing to remember is the inevitable trade-off between the two types of error.
By increasing the confidence level, we reduce the danger of the type-I error, but only
at the cost of increasing the danger of the type-II error; and vice versa.
• Explain how to evaluate statistically the results of an experimental measurement
of a classifier’s performance on a testing set.
• What is meant by the term, type-I error (false alarm)? What can be done to reduce
the danger of making this error?
• What is meant by the term, type-II error (missed detection)? What can be done
to reduce the danger of making this error?
12.5 Another Kind of Statistical Evaluation
At this moment, the reader understands the essence of statistical processing of

experimental results, and knows how to use it when evaluating the claims about
a given classifier’s performance. However, much more can be accomplished with
the help of statistics.
Do Two Testing Sets Represent Two Different Contexts? Chapter 10 mentioned
the circumstance that, sometimes, a different classifier should perhaps be induced
for a different context—such as the British accent as compared to the American
accent. Here is how statistics can help us identify such situations in the data.
Suppose we have tested two classifiers on two different testing sets. The
classification accuracy in the first test is pO 1 and the classification accuracy in the
second test is pO 2 (the letter “p” alluding to the proportion of correct answers). The
sizes of the two sets are denoted by n1 and n2 . Finally, let the average proportion of
correctly classified examples in the two sets combined be denoted by pO .
The statistics of interest is defined by the following formula:
pO 1 pO 2
zD q (12.8)
pO .1 pO /. n11 C 1
n2
/
The result is compared to the critical value for the given confidence level—the
value can be found in Table 12.2.
A Concrete Example Suppose the classifier was evaluated on two testings sets
whose sizes are n1 D 100 and n2 D 200. Let the classification accuracies measured
on the two be pO 1 D 0:82 and pO 2 D 0:74, respectively, so that the average
classification accuracy on the two sets combined is pO D 0:77. The reader will easily
verify that the conditions for normal distribution are satisfied.
Plugging these values into Eq. (12.8), we obtain the following:
0:82 0:74
zD q D 1:6: (12.9)
1 1
0:77.1 0:77/. 100 C 200
/
Since this value is lower than the one given for the 95% confidence level in
Table 12.2, we conclude that the result is within the corresponding confidence
interval, and therefore accept that the two results are statistically indistinguishable.
To make sure you understand the topic, try to answer the following questions.
If needed, return to the appropriate place in the text.
12.6 Comparing Machine-Learning Techniques 245
• Why should we be concerned that a classifier is being applied to a wrong kind of

data?
• What formula is used here? How do you carry our the evaluation?
12.6 Comparing Machine-Learning Techniques
Sometimes, we need to know which of two alternative machine-learning techniques

is more appropriate for the given class-recognition problem. The usual methodology
is here somewhat different than in the previous sections, though it is built around
similar principles. Instead of lengthy theorizing, let us illustrate this kind of
evaluation on a simple example with concrete data.
Experimental Methodology As discussed in Chap. 11, one way to compare,
experimentally, two machine-learning algorithms is to use 5 2 crossvalidation.
The reader will recall that, in this method, the set of available pre-classified data is
divided into two equally sized subsets, T11 and T12 . First, the two machine-learning
techniques are used to induce their classifiers from T11 , and these classifiers are then
tested on T12 ; then, the two classifiers are both induced from T12 and tested on T11 .
The process is repeated five times, each time for a different random division of the
set of data into two subsets, Ti1 and Ti2 .
As a result, we obtain ten pairs of testing-set classification accuracies (or error
rates, precisions, recalls, or any other performance criterion of choice). The question
to ask is then formulated as follows: “Are the differences between the ten pairs of
results statistically significant?”
Example of Experimental Results: Paired Comparisons Let us denote the i-th
pair of the sets by Ti1 and Ti2 , respectively. Suppose we are comparing two machine-
learning algorithms, ML1 and ML2, evaluating them in the ten experimental runs
as explained in the previous paragraph. Suppose that the results are those listed in
Table 12.4. In this table, each column is headed with the name of the test set used
in the corresponding experimental run. The fields in the table give the classification
accuracies (in percentages) that have been achieved on the corresponding test sets
by classifiers induced by the two alternative induction techniques. The last row
specifies the differences between the two classification accuracies. Note that the
differences are either positive or negative.
Evaluating these results, we realize that the average difference is d D 2:0, and
that the standard deviation of these differences is sd D 4:63.
The Principle of Statistical Evaluation of Paired Differences Observing the
mean value of differences, d (with standard deviation, sd ), we have to ask: is this
difference statistically significant? In other words, is this difference outside what
we previously called a confidence interval for a given confidence level, say, 95%?
Note that the midpoint of this confidence interval is dO D 0.
Table 12.4 Example experimental results of a comparison of two alternative machine-learning

techniques, ML1 and ML2
T11 T12 T21 T22 T31 T32 T41 T42 T51 T52
ML1 78 82 99 85 80 95 87 57 69 73
ML2 72 79 95 80 80 88 90 50 73 78
d 6 3 4 5 0 7 3 7 4 5
The numbers in the first two rows give classification accuracies (in percentages), the last row gives
the differences, d
Table 12.5 Some Degrees Confidence level

probabilities of the t-values
of freedom 0.10% 0.05% 0.02% 0.0%1
for nine degrees of freedom
9 1.83 2.26 2.81 3.35
As compared to the methods discussed in the previous sections, we have to point

out two major differences. First of them is the fact that, instead of proportions, we
are now dealing with mean values, d. The second is the circumstance that we can
no longer rely on the normal distribution because the number, n, of the values on
which the statistical evaluation rests is small, and the standard deviation has only
been estimated on the basis of the given ten observations (it is not known for the
entire population).
In this situation, we have to resort to another theoretical distribution, the so-called
t-distribution. Its shape is similar to the normal distribution (the “bell” shape), but
it is flatter. Also, its “flatness” or “steepness” depends on what is called the number
of degrees of freedom. In the case of 10 testing sets, there are 10 1 D 9 degrees
of freedom. Some typical t values for the case of 9 degrees of freedom are shown in
Table 12.5.3
Calculating the t-Values in Paired Tests Statistical evaluation using the t-tests is
essentially the same as in the case of normal distribution. For the mean difference,
d, and the standard deviation sd , the t9 -value (the subscript refers to the number of
degrees of freedom) is calculated by the following formula, where n is the number
of tests:
d0
t9 D p (12.10)
sd = n
The result is then compared with the critical thresholds associated with concrete
levels of confidence. These are listed in Table 12.5. Specifically, in the case of the
results from Table 12.4, we obtain the following:
3
With more degrees of freedom, the curve would get closer to the normal distribution, becoming
almost indistinguishable from it for 30 or more degrees of freedom.
20
t9 D p D 1:35 (12.11)
4:63= 10
Seeing that the obtained value is less than the 2.26 listed for the 95% confidence
level in the table, we conclude that the experiment has failed to refute (for the
given confidence level) the hypothesis that the two techniques lead to comparable
classification accuracies. We therefore accept the claim.
• Explain the principle of the 5 2-crossvalidation technique which results in a set
of 10 paired results.
• Why cannot we use the normal distribution as in the previous sections? What
other distribution is used here?
• Write down the formula calculating the t-value. Explain how the value for each
individual variable in this formula is obtained.
• The essence of statistical evaluation is to draw conclusions about the behavior of

an entire population based on our observations made on a relatively small sample.
• Different samples will yield different results, but these different values are bound
to be distributed according to statistical laws. Knowledge of these laws helps a
machine-learning engineer to calculate his or her confidence in the classification
performance as measured on concrete testing set.
• The most typical distribution of the “sampling results” is the Gaussian normal
distribution. It can only be used if two essential conditions are satisfied. For
training-set size n and for the mean value p, the conditions are as follows:
np 10 and n.1 p/ 10.
• Regardless whether the distribution is normal or not, the standard error of the
classification accuracies, acc, measured on different testing sets is given by the
following formula:
r r
acc.1 acc/ 0:75 0:25
sacc D D D 0:043
n 100
• For each confidence level, the normal-distribution assumption leads to a specific
z value (see Table 12.2). Having calculated the standard error, and having chosen
a confidence level, we establish the confidence interval by the following formula:
Œacc z sacc ; acc C z sacc
The term z sacc is referred to as the margin of error. For different performance
metrics, similar formulas are used.
• Suppose we are testing a claim about certain classification accuracy, acc. If the
result of experimental evaluation falls into the confidence interval defined by the
chosen confidence level, we assume we do not have enough evidence against
the claim regarding the value of acc. If the result is outside the value, we reject
the claim.
• When comparing two machine-learning techniques, we often rely on the 5 2-
crossvalidation technique, subjecting the results to the so-called t-test. This test
relies on the t-distribution instead of the normal distribution. The t-distribution
has a slightly different shape for different numbers of degrees of freedom.
Historical Remarks The statistical methods discussed in this chapter are so old
and well established that textbooks of statistics no longer care to give credit to those
who developed them. From the perspective of machine learning, however, we must
mention that the idea of applying t-tests to experimental results obtained from 5 2
crossvalidation was advocated by Dietterich [22].
Exercises
1. Suppose we intend to evaluate the statement that a certain classifier’s accuracy is

p D 0:80. What size n, of the testing set is needed if we want to be able to rely
on the normal distribution and Table 12.2?
2. Suppose that a certain classifier accuracy has been specified as p D 0:9. What is
the value of the standard error, sE , if the classifier is to be evaluated on testing
sets of size n D 400? Determine the size of the confidence interval for the
95% confidence level. Do not forget to check the validity of Conditions (12.1)
and (12.2).
3. Suppose your company is offered a classifier with stated accuracy p D 0:85.
When the company decides to test the validity of this statement on a testing set
of 200 examples, the measured value is 0.81, which of course is less than what
was promised. Is there at least 95% chance that the original claim was correct?
What about 99%?
1. Suppose you test a classifier’s performance, using the 95%-confidence interval.

What if you change your mind and decide to use the 99%-confidence instead?
You will increase tolerance, but what is the price for this?
1. This assignment assumes that the reader has already implemented a program
dividing the data into the five “folds” needed for the evaluation of the perfor-
mance using the 5 2-CV methodology. Another assumption is that the reader
has implemented at least two class-induction programs.
Write a program comparing the two induction techniques using the 5 2-CV
methodology, evaluating the results using t-tests.
Chapter 13
Induction in Multi-Label Domains
All the techniques discussed in the previous chapters assumed that each example
is labeled with one and only one class. In realistic applications, however, this is
not always the case. Quite often, an example is known to belong to two or more
classes at the same time, sometimes to many classes. For machine learning, this
poses certain new problems. After a brief discussion of how to deal with this issue
within the framework of classical paradigms, this chapter describes the currently
most popular approach: binary relevance.
The idea is to induce a binary classifier separately for each class, and then to
use all these classifiers in parallel. More advanced versions of this technique seek
to improve classification performance by exploiting mutual interrelations between
classes. As yet another alternative, the chapter discusses also the simple mechanism
of class aggregation.
13.1 Classical Machine Learning in Multi-Label Domains
Let us begin with an informal definition of a multi-label domain. After this, we will
take a look at how to address the problem within the classical paradigms introduced
in earlier chapters.
What is a Multi-Label Domain? In many applications, the traditional requirement
that an example should be labeled with one and only one class is hard to satisfy.
Thus a text document may represent nutrition, diet, athletics, popular
science, and perhaps quite a few other categories. Alternatively, an image may at
the same time represent summer, cloudy weather, beach, sea, seagulls,
and so on. Something similar is the case in many other domains.
The number of classes with which an average example is labeled differs from
one application to another. In some of them, almost every example has a great many
labels selected from perhaps thousands of different classes. At the other end of the

DOI 10.1007/978-3-319-63913-0_13
252 13 Induction in Multi-Label Domains
spectrum, we find domains where only some examples belong to more than one
class, the majority being labeled with only a single one.
Whatever the characteristics of the concrete data, the task for machine learning is
to induce a classifier (or a set of classifiers) satisfying two basic requirements. First,
the tool should for a given example return as many of its true classes as possible;
missing any one of them would constitute a false negative. At the same time, the
classifier should not label the example with a class to which the example does not
belong—each such “wrong” class would constitute a false positive.
Neural Networks Chapter 5 explained the essence of a multilayer perception,
MLP, a popular architecture of artificial neural networks. The reader will recall
that the output layer consists of one neuron for each class, the number of inputs
equals the number of attributes, and the ideal size of the hidden layer reflects the
complexity of the classification problem at hand.
On the face of it, using an MLP in multi-label domains should not pose any major
problems. For instance, suppose the network has been presented with a training
example that is labeled with classes C3 ; C6 ; and C7 . In this event, the target values
for training will be set to, say, ti D 0:8, in the case of output neurons with indices
i 2 f3; 6; 7g, and to ti D 0:2 for all the other output neurons.1 The backpropagation-
of-error technique can then be used in the same manner as in single-label domains.
A Word of Caution Multilayer perceptions may not necessarily be the best choice
here. Indeed, multi-label domains have been less intensively studied, in the neural-
networks literature, than other approaches, and not without reason. For one thing,
the training of plain MLPs is known to be vulnerable to local minima, and there is
always the architecture-related question: what is the best number of hidden neurons
if we want to strike a reasonable compromise between overfitting the data if the
network is too large, and suffering from insufficient flexibility if the network is too
small?
Also the notoriously high computational costs can be a reason for concern.
The fact that each training example can belong to more than one class certainly
complicates the learning process. Sensing the difficulty of the task, the engineer is
sometimes tempted to increase the number of hidden neurons. This, however, not
only adds to the already high computational costs, but also increases the danger of
overfitting.
It is always good to keep in mind that training neural networks is more art than
science. While a lot can be achieved through ingenuity and experience, beginners
are often disappointed. And in the case of a failure, the machine-learning expert
should be prepared to resort to some alternative, less dangerous technique.
1
The reader will recall that the target values 0:8 and 0:2 are more appropriate for the
backpropagation-of-error algorithm than 1 and 0. See Chap. 5.
13.1 Classical Machine Learning in Multi-Label Domains 253
Nearest-Neighbor Classifiers Another possibility is the use of nearest-neighbor

classifiers with which we got acquainted in Chap. 3. When example x is presented,
the k-NN classifier first identifies the example’s k nearest neighbors. Each of these
may have been labeled with a set of classes, and the simplest classification attempt in
a multi-label domain will label x with the union of these sets. For instance, suppose
that k D 3, and suppose that the sets of class labels encountered in the three nearest
neighbors are fC1 ; C2 g, fC2 g, and fC1 ; C3 g, respectively. In this event, the classifier
will classify x as belonging to C1 ; C2 , and C3 .
A Word of Caution This approach is practical only in domains where the average
number of classes per example is moderate, say, less than three. Also the number
of voting neighbors, k, should be small. Unless these two requirements are satisfied,
too many class labels may be returned for x, and this can give rise to too many
false positives, which, in turn, leads to poor precision. At the same time, however,
the multitude of returned labels also reduces the number of false negatives, which
improves recall. In some domains, this is what we want. In others, precision is
critical, and its low value may not be acceptable.
As so often in this paradigm, the engineer must resist the temptation to increase
the number of the nearest neighbors in the hope that spreading the vote over
more “participants” will give a chance to less frequent classes. The thing is, some of
these “nearest neighbors” might then be too distant from x, and thus inappropriate
for classification purposes.
A Note on Other Approaches Machine learning scientists have developed quite a
few other ways of modifying traditional machine-learning paradigms for the needs
of multi-label domains. Among these, very interesting appear to be attempts to
induce multi-label decision trees. But since they are somewhat too advanced for an
introductory text, we will not present them here. After all, comparable classification
performance can be achieved by simpler means—and these will be the subject of
the rest of this chapter.
• Suggest an example of a multi-label domain. What is the essence of the
underlying machine-learning task? In what way will one multi-label domain
differ from another?
• Explain the simple method of multi-label training in multilayer perceptions.
What practical difficulties might discourage you from using this paradigm?
• Describe the simple way of addressing a multi-label domain by a k-NN classifier.
Discuss its potential pitfalls.
13.2 Treating Each Class Separately: Binary Relevance
Let us now proceed to the main topic of this section, the technique of binary
relevance. We will begin by explaining the principle, and then discuss some of its
shortcomings and limitations.
The Principle of Binary Relevance The most common approach to multi-label
domains induces a separate binary classifier for each class: in a domain with N
classes, N classifiers are induced. When classifying a future example, all these
classifiers are used in parallel, and the example receives all classes for which the
classifiers returned the positive label.
For the induction of these classifiers, the training data have to be modified
accordingly. Here is how. For the i-th class (i 2 Œ1; N), we create a training set, Ti ,
that consists of the same examples as the original training set, T, the only difference
being in labeling: in Ti , an example’s class label is 1 if the list of class labels for this
example in T contains Ci ; otherwise, the label in Ti is 0.
Once the new training sets have been created, we apply to each of them a baseline
learner that is responsible for the induction of the individual classifiers. Common
practice applies the same baseline learner to each Ti . Typically, we use to this end
some of the previously discussed machine-learning techniques such as perception
learning, decision-tree induction, and so on.
Illustration of the Learning Principle Table 13.1 illustrates the mechanism with
which the new training data are created. In the original training set, T, five different
class labels can be found: C1 ; : : : ; C5 . The binary relevance technique creates the
five new training sets, T1 ; : : : ; T5 , shown in the five tables below the original one.
Table 13.1 The original multi-label training set is converted into five new training sets, one for
each class
Classes
ex1 C1 , C2
ex2 C2
ex3 C1 , C3 , C5
ex4 C2 , C3
ex5 C2 , C4
T1 T2 T3 T4 T5
ex1 1 ex1 1 ex1 0 ex1 0 ex1 0
ex2 0 ex2 1 ex2 0 ex2 0 ex2 0
ex3 1 ex3 0 ex3 1 ex3 0 ex3 1
ex4 0 ex4 1 ex4 1 ex4 0 ex4 0
ex5 0 ex5 1 ex5 0 ex5 1 ex5 0
13.2 Treating Each Class Separately: Binary Relevance 255
Thus in the very first of them, T1 , examples ex1 and ex3 are labeled with 1 because
these (and only these) two examples contain the label C1 in the original T. The
remaining examples are labeled with 0.
The baseline learner is applied separately to each of the five new sets, inducing
from each Ti the corresponding classifier Ci .
An Easy-to-Overlook Pitfall In each of the training sets thus obtained, every
example is labeled as a positive or negative representative of the given class. When
the induced binary classifiers are used in parallel (to classify some x), it may happen
that none of them returns 1. This means that no label for x has been identified.
When writing the machine-learning software, we must not forget to instruct the
classifier what to do in this event. Usually, the programmer chooses from the
following two alternatives: (1) return a default class, perhaps the one most frequently
encountered in T, or (2) reject the example as too ambiguous to be classified.
Discussion The thing to remember is that the idea behind binary relevance is
to transform the multi-label problem into a set of single-label tasks that are then
addressed by classical machine learning. To avoid disappointment, however, the
engineer needs to be aware of certain difficulties which, unless properly addressed,
may lead to underperformance. Let us briefly address them.
Problem 1: Imbalanced Classes Some of the new training sets, Ti , are likely to
suffer from the problem of imbalanced class representation which was discussed
in Sect. 10.2. In Table 13.1, this occurs in the case of sets T4 and T5 . In each of
them, only one example out of five (20%) is labeled as positive, and all others are
labeled as negative. In situations of this kind, we already know, machine-learning
techniques tend to be biased toward the majority class—in this particular case, the
class labeled as 0.
The solution is not difficult to find. The two most straightforward approaches
are majority-class undersampling or minority-class oversampling. Which of them
to choose will of course depend on the domain’s concrete circumstances. As a
rule of thumb, one can base the decision on the size of the training set. In very
big domains, majority-class undersampling is better; but when the examples are
scarce, the engineer cannot afford to “squander” them, and thus prefers minority-
class oversampling.
Problem 2: Computational Costs Some multi-label domains are very large.
Thus the training set in a text categorization domain may consist of hundreds of
thousands of examples, each described by tens of thousands of attributes and labeled
with a subset of thousands of different classes. It stands to reason that to induce
thousands of decision trees from a training set of this size will be expensive, perhaps
prohibitively so. We can see that, when considering candidates for the baseline
learner, we may have to reject some of them because of computational costs.
Another possibility is to resort to the technique discussed in Sect. 9.5 in the
context of boosting techniques: for each class, we create multiple subsets of the
training examples, some of them perhaps described by different subsets of attributes.
The idea is to induce for each class a group of subclassifiers that then vote. If (in
a given paradigm) learning from 50% of the examples takes only 5% of the time,
considerable savings can be achieved.
Problem 3: Performance Evaluation Another question is how to measure the
success or failure of the induced classifiers. Usually, each of them will exhibit
different performance, some better than average, some worse than average, and
some dismal. To get an idea of the big picture, some averaging of the results is
needed. We will return to this issue in Sect. 13.7.
Problem 4: Neglecting Mutual Interdependence of Classes The baseline version
of binary relevance treats all classes as if they were independent of each other.
Quite often, this assumption is justified. In other domains, the classes are to some
degree interdependent, but not much harm is done when this fact is ignored. But in
some applications, the overall performance of the induced classifiers considerably
improves if we find a way to exploit the class interdependence. To introduce
methods of doing so will be the task for the next three sections.
• Describe the principle of binary relevance. How does it organize the learning
process, and how are the induced classifiers used for the classification of future
examples?
• What can render the computational costs of this approach prohibitively high?
How will the engineer handle the situation?
• Why does binary relevance often lead to the problem of imbalanced classes?
What remedies would you recommend?
13.3 Classifier Chains
In many applications, the classes are interrelated. The fact that a text document
has been labeled as nutrition is sure to increase its chances of belonging also
to diet—and decrease the probability that has something to do with quantum
mechanics. In the context of binary relevance, this means that methods of
exploiting class interdependence are likely to improve classification performance.
One possibility of doing so is known as a classifier chain.
The Idea A very simple approach relies on a chain of classifiers such as the one
in Fig. 13.1. Here, the classifiers are created one at a time, starting from the left. To
begin with, the leftmost classifier is induced from the original examples labeled as
positive or negative instances of class C1 (recall the training set T1 from Table 13.1).
13.3 Classifier Chains 257
Fig. 13.1 With the exception of C1, the input of each classifier consists of the original attribute
vector plus the label returned by the previous classifier
The second classifier is then induced from examples labeled as positive or negative
instances of class C2 . To describe these latter examples, however, one extra attribute
is added to the original attribute vector: the output of C1 . The same principle is then
repeated in the course of the induction of all the remaining classifiers: for each, the
training examples are described by the original attribute vector plus the class label
returned by the previous classifier.
When using the classifier chain for the classification of some future example,
x, the same pattern is followed. The leftmost classifier receives x described by
the original attributes. To all other classifiers, the system presents x described
by the original attribute vector plus the label delivered by the previous classifier.
Ultimately, x is labeled with those classes whose classifiers returned 1.
An Important Assumption (Rarely Satisfied) In the classifier-chain technique,
the ordering of the classes from left to right is the responsibility of the engineer.
In some applications, this is easy because the classes form a logical sequence. Thus
in document classification, science subsumes physics, which in turn subsumes
quantum mechanics, and so on. If a document does not belong to science, it
is unlikely to belong to physics, either; it thus makes sense to choose science
as the leftmost node in the graph in Fig. 13.1, and to place physics next to it.
In other applications, class subsumption is not so obvious, but the sequence
can still be used without impairing the overall performance. Even when the
subsumptions are only intuitive, the engineer may always resort to a sequence
backed by experiments: she can suggest a few alternative versions, test them, and
then choose the one with the best results. Another possibility is to apply the classifier
chain only to some of the classes (where the interrelations are known), treating the
others as if only plain binary relevance was to be employed.
Hierarchically Ordered Classes Class interrelation does not have to be linear. It
can acquire forms that can only be reflected by a more sophisticated data structure,
perhaps a graph. In that case, we will need more advanced techniques such as the
one described in Sect. 13.5.
One Shortcoming of the Classifier-Chain Approach More often than not, the
engineer lacks any a priori knowledge about class interrelations. If she then still
wants to employ classifier chains, the best she can do is to create the classifier
sequence randomly. Of course, such ad hoc method cannot be guaranteed to work;
to insist on an inappropriate classifier sequence may be harmful to the point where
the classification performance of the induced system may fail to reach even that of
plain binary relevance.
Sometimes, however, there is a way out. If the number of classes is manage-

able (say, five), the engineer may choose to experiment with several alternative
sequences, and then choose the best one. But if the number of classes is greater,
the necessity to try many alternatives will be impractical.
Error Propagation The fact that the classifiers are forced into a linear sequence
makes them vulnerable to a phenomenon known as error propagation. Here is what
it means. When a classifier misclassifies an example, the incorrect class label is
passed on to the next classifier that uses this label as an additional attribute. An
incorrect value of this additional attribute may then sway the next classifier to a
wrong decision, which, too, is passed on, down the chain. In other words, an error
of a single class may result in additional errors being made by subsequent classifiers.
In this event, the classifier chain is likely to underperform. The thing to remember
is that the overall error rate strongly depends on the quality of the earlier classifiers
in the sequence.
One Last Comment The error-propagation phenomenon is less damaging if the
classifiers do not strictly return either 0 or 1. Thus a Bayesian classifier calculates
for each class its probability, a number from the interval Œ0; 1. Propagating this
probability through the chain is less harmful than the strict pos or neg. Similar
considerations apply to some other classifiers such as those from the paradigm
of neural networks (for instance, multilayer perceptions and radial-basis function
networks).
• Discuss the typical impact that class interdependence can have on the perfor-
mance of the binary relevance technique.
• Explain the principle of classifier chain. What can you say about the need to find
a proper sequence of classifiers?
• Explain the problem of error propagation in classifier chains. Is there anything
else to criticize about this approach?
13.4 Another Possibility: Stacking
In the light of the aforementioned drawbacks of classifier chains, some less risky
alternative is needed. One possibility is to rely on stacking.
Architecture and Induction The essence is illustrated in Fig. 13.2. Here, the
classifiers are arranged in two layers. The upper one represents plain binary
13.4 Another Possibility: Stacking 259
Fig. 13.2 The stacking

technique. The upper-layer
classifiers use as input the
original attribute vector. For C1 C2 C3 C4
the lower-layer classifiers,
this vector is extended by the
vector of the class labels
returned by the upper layer
C1’ C2’ C3’ C4’
relevance (independently induced binary classifiers, one for each class). More
interesting is the bottom layer. Here, the classifiers are induced from the training
sets where the original attribute vectors are extended by the list of the class labels
returned by the upper-layer classifiers. In the concrete case depicted in Fig. 13.2,
each attribute vector is preceded by four new binary attributes (because there are
four classes): the i-th attribute has value 1 if the i-th classifier in the upper layer has
labeled the example as belonging to class i; otherwise, this attribute’s value is 0.
Classification When the class labels of some future example x are needed, x
is presented first to the upper-layer classifiers. After this, the obtained classes labels
are added at the front of x’s original attribute vector as N new binary attributes
(assuming there are N classes), and the newly described example is presented in
parallel to the lower-layer classifiers. Finally, x is labeled with the classes whose
lower-layer classifiers have returned 1.
The underlying philosophy rests on the intuition that the performance of classifier
Ci may improve if this classifier is informed about the “opinions” of the other
classifiers—about the other classes to which x belongs.
An Example Consider an example that is described by a vector of four attributes
with these values: x D fa; f ; r; zg. Suppose that the upper-layer classifiers return
the following labels: C1 D 1; C2 D 0; C3 D 1; c4 D 0. In this event, the
lower-layer classifiers are all presented with the following example description:
x D f1; 0; 1; 0; a; f ; r; zg.
The classification behaviors of the lower-level classifiers can differ from those in
the upper layer. For example, if the lower-layer classifiers return 1; 1; 1; 0, the overall
system will label x with C1 ; C2 , and C3 , ignoring the original recommendations of
the upper layer.
Some Comments Intuitively, this approach is more flexible than classifier chains
because stacking makes it possible for any class to influence the recognition of any
other class. The engineer does not provide any a priori information about class
interdependence, assuming that the class interdependence (or the lack thereof) is
likely to be discovered in the course of the learning process.
When treated dogmatically, however, this principle may do more harm than good.
The fact that x belongs to Ci often has nothing to do with x belonging to Cj .
If this is the case, forcing the dependence link between the two (as in Fig. 13.2)
will be counterproductive. If most classes are mutually independent, the upper
layer may actually exhibit better classification performance than the lower layer,
simply because the newly added attributes (the classes obtained from the upper
layer) are irrelevant—and we know that irrelevant attributes can impair the results
of induction.
Proper understanding of this issue will guide the choice of the baseline learner.
Some approaches, such as induction of decision trees, or WINNOW are capable of
eliminating irrelevant attributes, thus mitigating the problem.
• How are interdependent classes addressed by the stacking approach? Discuss
both the induction phase and the classification phase.
• In what situations will stacking outperform binary relevance and/or a classifier
chain?
• Under what circumstances will you prefer binary relevance to stacking?
13.5 A Note on Hierarchically Ordered Classes
In some domains, the known class interdependence is more complicated than in the
cases we have considered so far. Our machine-learning techniques then have to be
modified accordingly.
An Example Figure 13.3 shows a small part of a class hierarchy that could have
been suggested by a specialist preparing the data for machine learning. Each node
in the graph represents one class.
The domain deals with the classification of text documents. The hierarchy is
interpreted in a way reminiscent of decision trees. To begin with, some documents
may belong to the class machine learning. A solid line emanating from the
corresponding node represents “yes,” and the dashed line represents “no.” Among
those documents that do belong to machine-learning, some deal with the topic
decision tree, others with k-NN classifiers, and so on (for simplicity,
most subclasses are omitted here).
13.5 A Note on Hierarchically Ordered Classes 261
machine
learning
decision k−NN
trees classif.
irrel.
pruning
attrib.
Fig. 13.3 Sometimes, the classes are hierarchically organized in a way known to the engineer in
advance
In the picture, the relations are represented by arrows that point from parent nodes
to child nodes. A node can have more than one parent, but a well-defined class
hierarchy must avoid loops. The data structure defining class relations of this kind is
known as a directed acyclic graph. In some applications, each node (except for the
root node) has one and only one parent. This more constrained structure is known
as a generalization tree.
Induction in Domains of This Kind Induction of hierarchically ordered classes is
organized in a way similar to binary relevance. For each node, the corresponding
training set is constructed, and from this training set, the baseline learner induces a
classifier. By doing so, the most common approach proceeds in a top-down manner
where the output of the parent class instructs the choice of the examples for the
induction of a child class.
Here is a way to carry this out in a domain where the classes are organized by a
generalization tree. First, the entire original training set is used for the induction of
the class located at the root of the tree. Next, the training set is divided into two parts,
one containing training examples that belong to the root class, the other containing
training examples that do not belong to this class. The lower-level classes are then
induced only from the relevant training sets.
A Concrete Example In the problem from Fig. 13.3, the first step is to induce a
classifier for the class machine learning. Suppose that the original training set
consists of the seven examples shown in Table 13.2. The labels of those examples are
then used to decide which examples to include in the training sets for the induction
of the child classes. For instance, note that only positive examples of machine
learning are included in the training sets for decision trees and k-NN
classifiers. Conversely, only negative examples of machine learning
are included in the training set for the induction of programming.
Two Major Difficulties to Be Aware Of The induction process is not as simple
as it looks. The first problem complicating the task is, again, the phenomenon of
error propagation. Suppose an example represents a text document from the field
Table 13.2 Illustration of a domain with hierarchically ordered classes
Machine learning
ex1 pos
ex2 pos
ex3 pos
ex4 pos
ex5 neg
ex6 neg
ex7 neg
Decision trees k-NN Programming

ex1 1 ex1 0 ex5 1
ex2 1 ex2 0 ex6 0
ex3 0 ex3 1 ex7 0
ex4 0 ex4 1
In some lower-level classes, the training sets contain only those training examples for which the
parent classifier returned pos; in others, only those for which the parent classifier returned neg
circuit analysis. If this example is mistakenly classified as belonging to

machine-learning, the classifier, misled by this information, will pass it on to
the next classifiers, such as decision trees, thus potentially propagating the
error down to lower levels.2
Another complication is that the training sets associated with the individual nodes
in the hierarchy are almost always heavily imbalanced. Again, appropriate measures
have to be taken—usually undersampling or oversampling.
Where Does the Class Hierarchy Come From? In some rare applications, the
complete class hierarchy is available right from the start, having been created
manually by the customer who has the requisite background knowledge about the
concrete domain. This is the case of some well-known applications from the field of
text categorization.
Caution is needed, though. Customers are not infallible, and the hierarchies
they develop often miss important details. They may suffer from subjectivity—with
consequences similar to those explained when we discussed classifier chains. In
some domains, only parts of the hierarchy are known. In this event, the engineer
has to find a way of incorporating this partial knowledge in the binary relevance
framework discussed earlier.
2
The reader has noticed that the issue is similar to the one we have encountered in the section
dealing with classifier chains.
13.6 Aggregating the Classes 263
• Give an example of a domain where the individual classes are hierarchically
ordered.
• Explain the training-set dividing principle for the induction of hierarchically
ordered classifiers.
• What are the most commonly encountered difficulties in the induction of
hierarchically ordered classes?
13.6 Aggregating the Classes
The situation is simpler if there are only a few classes. For instance, the number
of all class-label combinations in a domain with three classes cannot exceed seven,
assuming that each example is labeled with at least one class.
Creating New Training Sets In a domain of this kind, a sufficiently large training
set is likely to contain a sufficient number of representatives for each class
combination, and this makes it reasonable to treat each such combination as a
separate class. The general approach is similar to those we have already seen: from
the original training set, T, new training sets, Ti , are created, and from each, a
classifier is induced by the baseline learner. However, do not forget that, in class
aggregation, each Ti represents one combination of class labels, say, C2 AND C4 .
When, in the future, some example x is to be classified, it is presented to all of
these classifiers in parallel.
Illustration of the Principle The principle is illustrated in Table 13.3. Here,
the total number of classes in the original training set, T, is three. Theoretically,
the total number of class combinations should be seven. In reality, only five of
these combinations are encountered in T because no example is labeled with C3
alone, and no example is labeled with all three classes simultaneously. We therefore
create five tables, each defining the training set for one class combination. Note that
this approach deals only with those class combinations that have been found in the
original training set. For instance, no future example will be labeled as belonging to
C3 alone. This may be seen as a limitation, and the engineer will have to find a way
to address this issue in a concrete application.
Classification The programmer must not forget to specify what exactly is to
be done in a situation where more than one of these “aggregated” classifiers
returns 1. In some machine-learning paradigms, say, a Bayesian classifier, this is
easy because the classifiers are capable of quantifying their confidence in the class
Table 13.3 In a domain with a manageable number of class-label combinations, it is often

possible to treat each combination as a separate class
Classes
ex1 C1 , C2
ex2 C2
ex3 C1 , C3
ex4 C2 , C3
ex5 C1
C1 C2 C1 AND C2 C1 AND C3 C2 AND C3

ex1 0 ex1 0 ex1 1 ex1 0 ex1 0
ex2 0 ex2 1 ex2 0 ex2 0 ex2 0
ex3 0 ex3 0 ex3 0 ex3 1 ex3 0
ex4 0 ex4 0 ex4 0 ex4 0 ex4 1
ex5 1 ex5 0 ex5 0 ex5 0 ex5 0
they recommend. If two or more classifiers return 1, the master classifier simply
chooses the one with the highest confidence.
The choice is more complicated in the case of classifiers that only return 1 or
0 without offering any information about their confidence in the given decision.
In principle, one may consider merging the sets of classes. For example, suppose
that, for some example x, two classifiers return 1, and that one of the classifiers is
associated with classes C1 ; C3 , and C4 , and the other is associated with classes C3
and C5 . In this event, x will be labeled with C1 ; C3 ; C4 , and C5 .
Note, however, that this may easily result in x being labeled with “too many”
classes. The reader already knows that this may give rise to many false positives,
and thus lead to low precision.
Alternative Ways of Aggregation In the approach illustrated in Table 13.3, the
leftmost table (the one headed by C1 ) contains only one positive label because there
is only one training example in T labeled solely with this class. If we want to avoid
having to deal with training sets that are so extremely imbalanced, we need a “trick”
that would improve the class representations in Ti ’s.
Here is one possibility. In Ti , we will label with 1 each example whose set of
class labels in the original T contains C1 . By doing so, we must not forget that ex1
will thus be labeled as positive also in the table headed with (C1 AND C2 ).
Similarly, we will label with 1 all subsets of the set of classes found in a given
training set. For instance, if an example is labeled with C1 , C3 , and C4 , we will label
it with 1 in all training sets that represent nonempty subsets of {C1 , C3 , C4 }. This, of
course, improves only training sets for relatively “small” combinations (combining,
say, only one or two classes). For larger combinations, the problem persists.
13.7 Criteria for Performance Evaluation 265
A solution of the last resort will aggregate the classes only if the given
combination is found in a sufficient percentage of the training examples. If the
combination is rare, the corresponding Ti is not created. Although this means that
the induced classifiers will not recognize a certain combination, this may not be
such big loss if the combination is rare.
Some Criticism Class aggregation is not a good idea in domains where the number
of class combinations is high, and the training set size is limited. If these two
conditions are not satisfied, some of the newly created sets, Ti , are likely to contain
no more than just a few positive examples, and as such will be ill-suited for machine
learning: the training sets will be so imbalanced that all attempts to improve the
situation by minority-class oversampling or majority-class undersampling are bound
to fail—for instance, this will happen when a class combination is represented by
just a single example.
As a rule of thumb, in domains with a great number of different class labels,
where many combinations occur only rarely and some do not occur at all, the
engineer will prefer plain binary relevance or some of its variations (chaining or
stacking). Class aggregation is then to be avoided.
• Describe the principle of class aggregation. Explain separately the induction
process and the way the induced classifiers are used to classify future examples.
• What possible variations on the class-aggregation theme do you know?
• What main shortcoming can render this approach impractical in many realistic
applications?
13.7 Criteria for Performance Evaluation
We have mentioned earlier that performance evaluation in multi-label domains

depends on averaging the results across classes. Let us introduce and briefly discuss
some commonly used ways of doing so.
Macro-Averaging The simplest approach, macro-averaging, finds the values of the
given criterion for each class separately, and then calculates their arithmetic average.
Let L be the total number of classes. Here are the formulas that calculate macro-
precision, macro-recall, and macro-F1 from the values of these quantities for the
individual classes:
1X
L
PrM D pri
L iD1
1X
L
ReM D rei (13.1)
L iD1
1X
L
F1M D F1i
L iD1
Macro-averaging is suitable in domains where each class has approximately

the same number of representatives. In some applications, this requirement is not
satisfied, but the engineer may still prefer macro-averaging if he or she considers
each class to be equally important, regardless of its proportional representation in
the training set.
Micro-Averaging In the other approach, micro-averaging, each class is weighed
according to its frequency in the given set of examples. In other words, the
performance is averaged over all examples. Let L be the total number of classes.
Here are the formulas for micro-precision, micro-recall, and micro-F1 :
PL
NTPi
Pr D PL iD1
iD1 .NTPi C NFPi /

PL
NTPi
Re D PL iD1
(13.2)
iD1 .NTPi C NFNi /
2 Pr Re

F1 D
Pr C Re

Note that F1 is calculated from micro-precision and micro-recall and not from
the observed classifications of the individual examples.
Micro-averaging is preferred in applications where the individual classes cannot
be treated equally. For instance, the engineer may reason that good performance on
dominant classes is not really compromised by poor performance on classes that are
too rare to be of any importance.
A Numeric Example Let us illustrate these formulas using Table 13.4. Here, we
can see five examples. For each of them, the middle column lists the correct class
labels, and the rightmost column gives the labels returned by the classifier. The
reader can see minor discrepancies in the sense that the classifier has missed some
classes (causing false negatives). For instance, this is the case of class C3 being
missed in example ex3 . At the same time, the classifier labels some examples with
incorrect class labels, which constitutes false positives. For instance, this is the case
of example ex1 being labeled with class C3 .
13.7 Criteria for Performance Evaluation 267
Table 13.4 Illustration of performance evaluation in multi-label domains
The following table gives, for five testing examples, the known class labels versus the class
labels returned by the classifier.
True Classifier’s
classes classes
ex1 C1 ; C2 C1 ; C2 ; C3 ;
ex2 C2 C2 ; C4 ;
ex3 C1 ; C3 ; C5 C1 ; C5 ;
ex4 C2 ; C3 C2 ; C3 ;
ex5 C2 ; C4 C2 ; C5 ;
Separately for each class, here are the values of true positives, false positives, and false
negatives. Next to them are the corresponding values for precision and recall, again separately
for each class.
2 2
NTP1 D2 NFP1 D0 NFN1 D0 Pr1 D 2C0
D1 Re1 D 2C0
D1
4 4
D1 Re2 D 4C0
D1
1 1
D 0:5 Re3 D 1C1
D 0:5
0 0
D0 Re4 D 0C1
D0
1 1
D 0:5 Re5 D 1C0
D1
This is how the macro-averages are calculated:
1C1C0:5C0C0:5
PrM D 5
D 0:6
1C1C0:5C0C1
ReM D 5
D 0:7
Here is how the micro-averages are calculated:
2C4C1C0C1
Pr D .2C0/C.4C0/C.1C1/C.0C1/C.1C1/
D 0:73
2C4C1C0C1
Re D .2C0/C.4C0/C.1C1/C.0C1/C.1C0/
D 0:8
These discrepancies are then reflected in the numbers of true positives, false
positives, and false negatives. These, in turn, make it possible to calculate for each
class its precision and recall. After this, the table shows the calculations of the
macro- and micro-averages of these two criteria.
Averaging the Performance over Examples So far, the true and false positive and
negative examples were counted across individual classes. However, in domains
where an average example belongs to a great many classes, it can make sense to
average over the individual examples.
The procedure is in principle the same as before. When comparing the true class
labels with those returned for each example by the classifier, we obtain the numbers
of true positives, false positives, and false negatives. From these, we easily obtain
the macro-averages and micro-averages. The only thing we must keep in mind is
that the average is not taken over the classes, but over examples—thus in macro-
averages, we divide the sum by the number of examples, not by the number of
classes.
• Give the formulas for macro-averaging of precision, recall, and F1 .
• Give the formulas for micro-averaging of precision, recall, and F1 . Discuss the
difference between macro-averaging and micro-averaging.
• What is meant by “averaging the performance over examples”?
• In some domains (such as text categorization), each example can be labeled with
more than one class at the same time. These are so-called multi-label domains.
• In domains of this kind, classical machine-learning paradigms can sometimes be
used. Unless special precautions have been taken, however, the results are rarely
encouraging. For some paradigms, multi-label versions exists, but these are too
advanced for an introductory text, especially in view of the fact that good results
can be achieved with simpler means.
• The most common approach to multi-label domains induces a binary classifier
for each class separately, and then submits the example to all these classifiers in
parallel. This is called the binary relevance technique.
• What the basic version of binary relevance seems to neglect is the fact that the
individual classes may not be independent of each other. The fact that an example
has been identified as a representative of class CA may strengthen or weaken its
chances of belonging also to class CB .
• The simplest mechanism for dealing with class interdependence in multi-label
domains is the classifier chain. Here, the output of one binary classifier is used as
an additional attribute describing the example to be presented to the next classifier
in line.
• One weakness of classifier chains is that the user is expected to specify the
sequence of classes (perhaps according to class subsumption). If the sequence
is poorly designed, the results are disappointing.
• Another shortcoming is known as error propagation: an incorrect label given
to an example by one classifier is passed on to the next classifier in the chain,
potentially misleading it.
• A safer approach relies on the two-layered stacking principle. The upper-layer
classifiers are induced from examples described by the original attribute vectors,
and the lower-layer classifiers are induced from examples described by attribute
vectors to which the class labels obtained in the upper layer have been added.
When classifying an example, the outcomes of the lower-layer classifiers are
used.
• Sometimes, it is possible to take advantage of known hierarchical order among
the classes. Here, too, induction is carried out based on specially designed
training sets. Again, the user has to be aware of the dangers of error propagation.
• Yet another possibility is to resort to class aggregation where each combination
of classes is treated as a separate higher-level class. A special auxiliary training
set is created for each of these higher-level classes.
• The engineer has to pay attention to ways of measuring the quality of the induced
classifiers. Observing that each class may experience different classification
performance, we need mechanisms for averaging over the classes (or examples).
Two of them are currently popular: micro-averaging and macro-averaging.
Historical Remarks The problem of multi-label classification is relatively new.
The first time it was encountered was in the field of text categorization—see McCal-
lum [57]. The simplest approach, the binary relevance principle, was employed by
Boutell et al. [7]. A successful application of classifier chains was reported by Read
et al. [79], whereas Goldpole and Sarawagi [32] are credited with having developed
the stacking approach. Apart from the approaches related to binary relevance, some
authors have studied ways of modifying classical single-label paradigms. The ideas
on nearest-neighbor classifiers in multi-label domains are borrowed from Zhang
and Zhou [101] (whose technique, however, is much more sophisticated than the
one described in this chapter). Induction of hierarchically ordered classes was first
addressed by Koller and Sahami [47]. Multi-label decision trees were developed by
Clare and King [15].
Table 13.5 An example of a True Classifier’s

multi-label domain classes classes
ex1 C1 C1 ; C2
ex2 C1 ; C2 C1 ; C2
ex3 C1 , C3 C1
ex4 C2 , C3 C2 , C3
ex5 C2 C2
ex6 C1 C1 ; C2
Exercises
1. Consider the multi-label training set shown in the left part of Table 13.5. Show
how the auxiliary training sets will be created when the principle of binary
relevance is to be used.
2. For the same training set, create the auxiliary training sets for the approach
known as class aggregation. How many such sets will we need?
3. Draw the schema showing how the problem from Table 13.5 would be addressed
by stacking. Suppose the examples in the original training set are described by
ten attributes. How many attributes will the lower-level classifiers have to use?
4. Suggest the classifier-chain schema for a domain with the following four
classes: decision trees, machine learning, classification,
pruning.
5. Returning to the set of examples from Table 13.5, suppose that a classifier has
labeled them as indicated in the rightmost column. Calculate the macro- and
micro-averages of precision and recall.
1. Suggest a multi-label domain where the principle of classifier chain can be a

reasonable strategy to follow. What would be the main requirement for such data?
2. Consider a domain where the majority of training examples are labeled each with
only a single class, and only a small subset of the examples (say, 5%) are labeled
with more than one class. Suggest a machine learning approach to induce reliable
classifiers from such data.
3. Suppose that you have a reason to assume that a few classes are marked by strong
interdependence while most of the remaining classes are mutually independent.
You are thinking of using the stacking approach. What is the main problem that
might compromise the performance of the induced classifiers? Can you suggest
a mechanism that overcomes this pitfall?
4. Suppose you have been asked to develop machine-learning software for induction
from multi-label examples. This chapter has described at least four approaches
that you can choose from. Write down the main thoughts that would guide your
decision.
5. Suggest a mechanism that would mitigate the problem of error propagation
during multi-label induction with hierarchically ordered classes. Hint: after a
testing run, consider “enriching” the training sets by “problematic” examples.
1. Write a program that accepts as input a training set of multi-label examples, and
returns as output the set of auxiliary training sets needed for the binary relevance
approach.
2. Write a program that converts the training set from the previous question into
auxiliary training sets, following the principle of class aggregation.
3. Search the web for machine-learning benchmark domains that contain multi-
label examples. Convert them using the data-processing program from the
previous question, and then induce the classifiers by the binary relevance
approach.
4. Write a program that first induces the classifiers using binary relevance as in the
previous question. In the next step, the program redescribes the training examples
by adding to their attribute vectors the class labels as required by the lower layer
in the classifier stacking technique.
5. What data structures would you use for the input and output data when
implementing the classifier stacking technique?
6. Write a program that takes as input the values of NTP ; NTN ; NFP ; NFN for each
class, and returns micro- and macro-averaged precision, recall, and F1 .
Chapter 14
Unsupervised Learning
It would be a mistake to think that machine learning always requires examples with
class labels. Far from it! Useful information can be gleaned even from examples
whose classes are not known. This is sometimes called unsupervised learning, in
contrast to the term supervised learning which is used when talking about induction
from pre-classified examples.
While supervised learning focuses on induction of classifiers, unsupervised
learning is interested in discovering useful properties of available data. Perhaps
the most popular task looks for groups (called clusters) of similar examples. The
centroids of these groups can then be used as gaussian centers for Bayesian or
RBF classifiers, as predictors of unknown attribute values, and even as visualization
tools for multidimensional data. Last but not least, techniques used in unsupervised
learning can be used to create higher-level attributes from existing ones.
The chapter describes some practical techniques for unsupervised learning,
explaining the basic algorithms, their behaviors in practical circumstances, and the
benefits they offer.
14.1 Cluster Analysis
The fundamental task in unsupervised learning is cluster analysis. Here, the input
is a set of examples, each described by a vector of attribute values—but no class
labels. The output is a set of two or more clusters of examples.
Identifying Groups of Similar Examples Figure 14.1 shows a simple domain
with a few examples described by two attributes: weight and height. An
observer can easily see that the examples form three or four groups, depending on
the subjective “level of resolution.”

DOI 10.1007/978-3-319-63913-0_14
274 14 Unsupervised Learning
Fig. 14.1 A two-dimensional
height
domain with clusters of
examples
weight
Visual identification of such groups in a two-dimensional space is easy, but in

four or more dimensions, humans can neither visualize the data nor see the clusters.
These can only be detected by cluster-analysis algorithms.
Representing Clusters by Centroids To begin with, we have to decide how
the clusters are to be described. A few alternatives can be considered: it is
possible to specify the clusters’ locations, sizes, boundaries, and perhaps some
other aspects. But the simplest approach relies on centroids.1 If all attributes are
numeric, the centroid is identified with the averages of the individual attributes.
For instance, suppose a two-dimensional cluster consists of the following examples:
.2; 5/; .1; 4/; .3; 6/. In this case, the centroid is described by vector .2; 5/ because
the first attribute’s average is 2C1C3 3
D 2 and the second attribute’s average is
5C4C6
3
D 5.
The averages can be calculated even when the attributes are discrete if we know
how to turn them into numeric ones. Here is a simple way of doing so. If the attribute
can acquire three or more different values, we can replace each attribute-value pair
with one boolean variable (say, season=fall, season=winter, etc.). The
values of the boolean attributes are then represented by 0 or 1 instead of false and
true, respectively.
What Should the Clusters Be Like? Clusters should not overlap each other: each
example must belong to one and only one cluster. Within the same cluster, the
examples should be relatively close to each other, certainly much closer than to
the examples from the other clusters.
An important question will ask how many clusters the data contain. In Fig. 14.1,
we noticed that the human observer discerns either three or four clusters. However,
the scope of existing options is not limited to these two possibilities. At one extreme,
the entire training set can be thought of as forming one big cluster; at the other,
1
Machine learning professionals sometimes avoid the term “center” which might imply mathemat-
ical properties that are for the specific needs of cluster analysis largely irrelevant.
14.1 Cluster Analysis 275
each example can be seen as representing its own single-example cluster. Practical
implementations often side-step the problem by asking the user to supply the
number of clusters by way of an input parameter. Sometimes, however, machine-
learning software is expected to determine the number automatically.
Problems with Measuring Distances Algorithms for cluster analysis usually need
a mechanism to evaluate the distance between an example and a cluster. If the cluster
is described by its centroid, the Euclidean distance between the two vectors seems
to offer a good way of doing so—assuming that we are aware of situations where
this can be misleading. This last statement calls for an explanation.
Euclidean distance may be inconvenient in the case of discrete attributes, but we
already know how to deal with them. More importantly, we must not forget that each
attribute is likely to represent a different quantity, which renders the use geometric
distance rather arbitrary: a 4-year difference in age is hard to compare with a 4-foot
difference in height. Also the problem of scaling plays its role: if we replace feet
with miles, the distances will change considerably.
We have encountered these problems earlier, in Chap. 3. In the context of cluster
analysis, however, these issues tend to be less serious than in k-NN classifiers. Most
of the time, engineers get around the difficulties by normalizing all attribute values
into the unit interval, xi 2 Œ0; 1. We will return to normalization in Sect. 14.2.
A More General Formula for Distances If the examples are described by
a mixture of numeric and discrete attributes, we can rely on the sum of the
squared distances along corresponding attributes. More specifically, the following
expression is recommended (denoting by n the number of attributes):
q
dM .x; y/ D †niD1 d.xi ; yi / (14.1)
In this formula, we will use d.xi ; yi / D .xi yi /2 for continuous attributes. For
discrete attributes (including boolean attributes), we put d.xi ; yi / D 0 if xi D yi and
d.xi ; yi / D 1 if xi ¤ yi .
Which Cluster Should an Example Belong To? Let us suppose that each example
is described by an attribute vector, x, and that each cluster is defined by its
centroid—which, too, is an attribute vector.
Suppose there are N clusters whose centroids are denoted by ci , where i 2 .1; N/.
The example x has a certain distance d.x; ci /, from each centroid. If d.x; cP / is the
smallest of these distances, it is natural to expect that x be placed in cluster cP .
For instance, suppose that we use the Euclidean distance, and that there are three
clusters. If the centroids are c1 D .3; 4/, c2 D .4; 6/ and c3 D .5; 7/, and if the
example is x Dp .4; 4/, then the Euclidean distances are d.x; c1 / D 1, d.x; c2 / D 2,
and d.x; c3 / D 10. Since d.x; c1 / is the smallest of the three values, we conclude
that x should belong to c1 .
Benefit 1: Estimating Missing Values The knowledge of the clusters can help us
estimate missing attribute values. Returning to Fig. 14.1, the reader will notice that
if weight is low, the example is bound to belong to the bottom-left cluster. In this
case, also height is likely to be low because it is low in all examples found in
this cluster. This aspect tends to be even more strongly pronounced in realistic data
described by multiple attributes.
In example descriptions, some attribute values are sometimes unknown. As a
simple way of dealing with this issue, Sect. 10.4 suggested that the missing value be
estimated as the average or the most frequent value encountered in the training set.
However, an estimate of the missing value as the average or the most frequent value
of the given cluster is sounder because it uses more information about the nature of
the domain.
Benefit 2: Reducing the Size of RBF Networks and Bayesian Classifiers Cluster
analysis can assist such techniques as Bayesian learners and radial-basis-function
networks. The reader will recall that these paradigms operate with centers.2 In the
simplest implementation, the centers are identified with the attribute vectors of the
individual examples. In domains with millions of examples, however, this would
lead to impractically big classifiers. The engineer then prefers to divide the training
set into N clusters, and to identify the gaussian centers with the centroids of the
clusters.
Benefit 3: A Simple Classifier Finally, the knowledge of data clusters may be
useful in supervised learning. It is quite common that all (or almost all) examples
in a cluster belong to the same class. In that case, the developer of a supervised-
learning software may decide first to identify the clusters, and then label each cluster
with its dominant class.
• What is the main difference between unsupervised learning and supervised
learning?
• What is a cluster of examples? Define some mechanisms to describe the clusters.
How can we measure the distance between an example and a cluster?
• Summarize the main benefits of cluster analysis.
2
For instance, the section on RBF networks denoted these centers by i ’s.
14.2 A Simple Algorithm: k-Means 277
14.2 A Simple Algorithm: k-Means
Perhaps the simplest algorithm to detect clusters of data is known under the name k-
means. The “k” in the name denotes the requested number of clusters—a parameter
whose value is supplied by the user.
The Outline of the Algorithm The pseudocode of the algorithm is provided in
Table 14.1. The first step creates k initial clusters such that each example finds itself
in one and only one cluster. After this, the coordinates of all centroids are calculated.
Let us note that the problem of initialization is somewhat more complicated than
that. We will return to this issue presently.
In the next step, k-means investigates one example at a time, calculating its
distances from all centroids. The nearest centroid then defines the cluster to which
the example should belong. If the example already is that cluster, nothing needs to be
done; otherwise, the example is transferred from the current (wrong) cluster to the
right one. After the relocation, the centroids of the two affected clusters (the one that
lost the example, and the one that gained it) have to be recalculated. The procedure
is graphically illustrated by the single-attribute domain from Fig. 14.2. Here, two
examples find themselves in the wrong cluster and are therefore relocated. Note
how, after the example relocation, the vertical bars (separating the clusters), and
also the centroids, change their locations.
Termination The good thing about the algorithm described in Table 14.1 is that
the process is guaranteed to reach a situation where each example finds itself in the
nearest cluster so that, from this moment on, no further transfers are needed. The
clusters do not overlap. Since this is usually achieved in a manageable number of
steps, no sophisticated termination criterion is needed here.
Table 14.1 The clustering algorithm k-means

Input: a set of examples without class labels
user-set constant k
1. Create k initial clusters. For each, calculate the coordinates of its centroid, Ci , as the
numeric averages of the attribute values in the examples it contains.
2. Choose an example, x, and find its distances from all centroids. Let j be the index of the
nearest centroid.
3. If x already finds itself in the j-th cluster, do nothing. Otherwise, move x from its current
cluster to the j-th cluster and recalculate the centroids.
4. Unless a stopping criterion has been satisfied, repeat the last two steps for another
example.
Stopping criterion: each training example already finds itself in the nearest cluster.
Let us consider an almost trivial domain where 13 examples are described by a single
numeric attribute. Suppose the examples have been initially divided into the three
groups indicated here by the vertical bars. The following sequence shows how two
examples (marked by circles) are moved from one cluster to another.
initial clustering in a wrong cluster
examples
+ + + centers
in a wrong cluster
+ + +
+ + +
now: all examples in correct clusters
After the second transfer, the clusters are perfect and the calculations can stop.
Fig. 14.2 Illustration of the k-means procedure in a domain with one numeric attribute
Numeric Example In Table 14.2, a set of nine two-dimensional examples has

been randomly divided into three groups (because the user specified k D 3), each
containing the same number of examples. The table also provides the centroids for
each group. k-means goes through these examples systematically, one by one—in
this concrete case, starting with group-2. For each example, its distance from each
centroid is calculated. It turns out that the first example from group-2 already finds
itself in the right cluster. However, the second example is closer to group-1 than to
group-2 and, for this reason, has to be transferred from its original cluster to group-
1. After this, the affected centroids are recalculated.
14.2 A Simple Algorithm: k-Means 279
Table 14.2 Illustration of the k-means procedure in a domain with two attributes
The table below contains three initial groups of vectors. The task is to find “ideal” clusters
using the k-means (k D 3).
Group-1 Group-2 Group-3

.2; 5/ .4; 3/ .1; 5/
.1; 4/ .3; 7/ .3; 1/
.3; 6/ .2; 2/ .2; 3/
Centroids: .2; 5/ .3; 4/ .2; 3/
Let us pick the first example in group-2. The Euclidean

p p distances p between this example,
.4; 3/, and the centroids of the three groups are 8, 2, and 4, respectively. This means
that the centroid of group-2 is the one nearest to the example. Since this is where the
example already is, k-means does not do anything.
Let
p uspnow proceed
p to the second example in group-2, .3; 7/. In this case, the distances are
5; 9, and 17, respectively. Since the centroid of group-1 has the smallest distance, the
example is moved from group-2 to group-1. After this, the averages of the two affected
groups are recalculated.
Here are the new clusters:
Group-1 Group-2 Group-3

.2; 5/ .4; 3/ .1; 5/
.1; 4/ .2; 2/ .3; 1/
.3; 6/ .2; 3/
.3; 7/
Averages: .2:25; 5:25/ .3; 2:5/ .2; 3/
The process continues as long as any example transfers are needed.
The Need for Normalization The reader will recall that, when discussing k-NN
classifiers, Chap. 3 argued that inappropriate attribute scaling will distort the
distances between attribute vectors. The same concern can be raised in cluster
analysis. It is therefore always a good idea to normalize the vectors so that all
numeric attributes have values from the same range, say, from 0 to 1.
The simplest way of doing so is to determine for the given attribute its maximum
(MAX) and minimum (MIN) value in the training set. Then, each value of this
attribute is re-calculated using the following formula:
x MIN
xD (14.2)
MAX MIN
As for boolean attributes, their values can simply be replaced with 1 and 0 for true
and false, respectively. Finally, an attribute that acquires n discrete values (such as
season, which has four different values) can be replaced with n boolean attributes,
one for each value—and, again, for the values of these boolean attributes, 1 or 0
are used.
Computational Aspects of Initialization To reach its goal, the k-means needs to
go through a certain number of transfers of examples from wrong clusters to the
right clusters. How many such transfers are needed depends on the contents of the
initial clusters. At least in theory, it can happen that the randomly created initial
clusters are already perfect, and not a single example needs to be moved. This,
of course, is an extreme, but the reader surely understands that if the initialization
is “lucky,” fewer relocations have to be carried out than otherwise. Initialization
matters in the sense that a better starting point ensures that the solution is found
sooner.
How to Initialize In some domains, we can take advantage of some background
knowledge about the problem at hand. For instance, seeking to create initial clusters
in a database of a company’s employees, the data analyst may speculate that it makes
sense to group them by their age, salary, or some other intuitive criterion, and that
the groups thus obtained will be good initial clusters.
In other applications, however, no such guidelines exist. The simplest procedure
then picks k random training examples and regards them as code vectors to
define initial centroids. The initial clusters are then created by associating each of
the examples with its nearest code vector.
A More Serious Problem with Initialization There is another issue, and a more
serious one than the computational costs. The thing is, also the composition of
the resulting clusters (once the k-means has completed its work) may depend on
initialization. Choose a different set of initial code vectors, and the technique may
generate a different set of clusters.
The point is illustrated in Fig. 14.3. Suppose that the user wants two clusters
(k D 2). If he chooses as code vectors the examples denoted by x and y then the
initial clusters created with the help of these two examples are already perfect.
Fig. 14.3 Suppose k D 2. If

height
the code vectors are [x,y], the
initial clusters for k-means X
will be different than when
the code vectors are [x,z] Z
weight
14.3 More Advanced Versions of k-Means 281
The situation changes when we choose for the code vectors examples x and z.
In this event, the two initial clusters will have a very different composition, and k-
means is likely to converge on a different set of clusters. The phenomenon will be
more pronounced if there are “outliers,” examples that do not apparently belong to
any of the two clusters.
Summary of the Main Problems The good thing about k-means is that it is easy to
explain and easy to implement. Yet this simplicity comes at a price. The technique
is sensitive to initialization; the user is expected to provide the number of clusters
(though he may not know how many clusters there are); and, as we will see, some
clusters can never be identified, in this manner. The next sections will take a look at
some techniques to overcome these shortcomings.
• Describe the principle of the k-means algorithm. What is the termination
criterion?
• What would be the consequence if we did not normalize the training set? Write
down the simple normalization formula.
• Describe some methods for the initialization of k-means. What are the main
consequences of good or bad initialization?
14.3 More Advanced Versions of k-Means
The previous section described the baseline version of k-means and explained its
main weaknesses: sensitivity to initialization, and lack of flexibility in deciding
about the number of clusters. Let us now take a look at some improvements meant
to address these shortcomings.
Quality of the Clusters An engineer always needs to be able to evaluate and
compare alternative solutions. And thus also for the topic addressed by this section,
we need a criterion to help us choose between different sets of clusters. Primarily,
the criterion should reflect the fact that we want clusters that minimize the average
distance between examples and the centroids of “their” clusters.
Let us denote by d.x; c/ the distance between example x and the centroid, c, of
the cluster to which x belongs. If all attributes are numeric, and if they all have been
normalized, then d.x; c/ can be evaluated either by the Euclidean distance or by the
more general Eq. (14.1).
The following formula (in which SD stands for summed distances) sums up the
.j/
distances of all examples from their clusters’ centroids. Here, xi denotes the i-th
example in the j-th cluster, K is the number of clusters, nj is the number of examples
in the j-th cluster, and cj is the j-th cluster’s centroid.
X
K X
nj
.j/
SD D d.xi ; cj / (14.3)
jD1 iD1
In cluster analysis, we seek to minimize SD. When calculating this quantity, we

must not forget that the value obtained by Eq. (14.3) will go down if we increase
the number of clusters (and thus decrease their average size), reaching SD D 0
in the extreme case when each cluster is identified with one and only one training
example. The formula is therefore useful only if we compare solutions that have
similar numbers of clusters.
Using Alternative Initializations Knowing that the composition of the result-
ing clusters depends on the algorithm’s initialization, we can suggest a simple
improvement. We will define two or more sets of initial code vectors, and apply
k-means separately to each of them. After this, we will evaluate the quality of all the
alternative data partitionings thus obtained, using the criterion defined by Eq. (14.3).
The best solution is the one for which we get the lowest value. This solution is then
retained, and the others discarded.
Experimenting with Different Values of k One obvious weakness of k-means
is the requirement that the user should provide the value of k. This is easier said
than done because, more often than not, the engineer has no idea into how many
clusters the available data naturally divide. Unless more sophisticated techniques
are used (about these, see later), the only way out is to try a few different values,
and then pick the best according to an appropriate criterion (such as the one defined
in Eq. (14.3)). As we already know, the shortcoming of this criterion is that it tends
to give preference to small clusters. For this reason, data analysts often normalize
the value of SD by k, the number of clusters.
Post-processing: Merging and Splitting Clusters The quality of the set of
clusters created by k-means can often be improved by post-processing techniques
that either increase the number of clusters by splitting, or decrease the number by
merging.
As for merging, two neighboring clusters will be merged if their mutual distance
is small. To find out whether the distance merits the merging, we simply calculate
the distance of two centroids, and then compare it with the average cluster-to-cluster
distance calculated by the following sum, where ci and cj are centroids:
X
SD d.ci ; cj / (14.4)
i¤j
14.4 Hierarchical Aggregation 283
Conversely, splitting makes sense when the average example-to-example dis-

tance within some cluster is high. The concrete solution is not easy to formalize
because once we have specified that cluster C is to be split into C1 and C2 , we need
to decide which of C’s examples will go to the first cluster and which to the second.
Very often, however, it is perfectly acceptable to identify in C two examples with the
greatest mutual distance, and then treat them as the code vectors of newly created
C1 and C2 , respectively.
Hierarchical Application of k-Means Another modification relies on recursive
calls. The technique begins with running k-means for k D 2, obtaining two clusters.
After this, k-means is applied to each of these two clusters separately, again with
k D 2. The process is continued until an appropriate termination criterion has been
satisfied— for instance, the maximum number of clusters the user wants to obtain,
or the minimum distance between neighboring clusters.
This hierarchical version side-steps one serious shortcoming of k-means: the
necessity to provide the number of clusters because, in this case, this is found
automatically. This is particularly useful when the goal is to identify the gaussian
centers in Bayesian classifiers or in RBF networks.
• Discuss some shortcomings of the k-means algorithm, and describe the simple
techniques for overcoming them.
• Specify how you would implement the cluster-splitting and cluster-merging
techniques described in this section.
• Explain the principle of hierarchical application of k-means.
14.4 Hierarchical Aggregation
Just as any other machine-learning technique, k-means has not only advantages, but
also shortcomings. To avoid at least some of the latter, we need an alternative, an
approach which is likely to prove useful in situations where k-means fails.
Another Serious Limitation of k-Means By minimizing the distance between the
examples and the centroids of the clusters to which these examples belong, k-means
essentially guarantees that it will discover “convex” clusters. Most of the time, this
is indeed what we want. Such clusters can be useful in the context of Bayesian
classifiers and the RBF networks where they reduce, sometimes very significantly,
the number of employed gaussian centers.
height
weight
Fig. 14.4 Note that the leftmost examples in the “bottom” cluster are closer to the “upper”
cluster’s centroid than to its own. In a domain of this kind k-means will not find the best solution
This approach, however, will do a poor job if the clusters are of a different
(“non-convex”) nature. To see the point, consider the clusters in Fig. 14.4. Here,
the leftmost example, x, in the bottom cluster is closer to the centroid of the upper
cluster, and k-means would therefore relocate it accordingly—and yet we feel that
this would not do justice to the nature of the two groups.
To deal with data of this kind, we need another technique, one capable of
identifying clusters such as those in Fig. 14.4.
An Alternative Way of Measuring Inter-Cluster Distance In the previous
section, the distance between two clusters was evaluated as the Euclidean distance
between their centroids. However, for the needs of the approach described below, we
will suggest another mechanism: we will measure the distances between all pairs of
examples, [x,y], such that x comes from the first cluster and y from the second. The
smallest value found among all these example-to-example distances then defines the
distance between the two clusters.
Returning to Fig. 14.4, the reader will agree that, along this new distance metric,
example x is closer to the bottom cluster than to the upper clusters. This means that
the limitation mentioned in the previous paragraph has been in this particular case
eliminated. The price for this improvement is increased computational costs: if NA
is the number of examples in the first cluster, and NB the number of examples in the
second, then NA NB example-to-example distances have to be evaluated. Most of
the time, however, the clusters are not going to so big as to make this an issue.
Numeric Example For the sake of illustration of how this new distance metric is
calculated, consider the following two clusters, A and B.
14.4 Hierarchical Aggregation 285
A B
x1 D .1; 0/ y1 D .3; 3/
x2 D .2; 2/ y2 D .4; 4/
Table 14.3 The basic algorithm of hierarchical aggregation

Input: a set of examples without labels
1. Let each example form one cluster. For N examples, this means creating N clusters, each
containing a single example.
2. Find a pair of clusters with the smallest cluster-to-cluster distance. Merge the two clusters
into one, thus reducing the total number of clusters to N 1.
3. Unless a termination criterion is satisfied, repeat the previous step.
Using the Euclidean formula, we calculate the individual example-to-example

distances as follows:
p
d.x1 ; y1 / D p13,
d.x1 ; y2 / D p25,
d.x2 ; y1 / D p2,
d.x2 ; y2 / D 8.
p
p2 ; y1 / D 2, we conclude that
Observing that the smallest of these values is d.x
the distance between the two clusters is d.A; B/ D 2.
Hierarchical Aggregation For domains such as the one in Fig. 14.4, the clustering
technique known as hierarchical aggregation is recommended. The principle is
summarized by the pseudocode in Table 14.3.
In the first step, each example defines its own cluster. This means that in a domain
with N examples, we have N initial clusters. In a series of the subsequent steps,
hierarchical aggregation always identifies a pair of clusters that have the smallest
mutual distance along the distance metric from the previous paragraphs. These
clusters are then merged. At an early stage, this typically amounts to merging pairs
of neighboring examples. Later, this results either in adding an example to its nearest
cluster, or in merging two neighboring clusters. The first few steps are illustrated
in Fig. 14.5.
The process continues until an appropriate termination criterion is satisfied. One
possibility is to stop when the number of clusters drops below a user-specified
threshold. Alternatively, one can base the stopping criterion on the cluster-to-cluster
distance (finish when the smallest of these distances exceeds a certain value).
height
height
x x
weight weight
Fig. 14.5 Hierarchical aggregation after first two steps (left) and after first nine steps (right). Note
how the clusters are gradually developed
• What kind of clusters cannot be detected by the k-means algorithm?

• What distance metric is used in hierarchical aggregation? What are the advan-
tages and disadvantages of this metric?
• Describe the principle of the hierarchical-aggregation approach to clustering. For
what kind of clusters is it particularly suited?
14.5 Self-Organizing Feature Maps: Introduction
Let us now introduce yet another approach to unsupervised learning, this time
borrowing from the field of neural networks. The technique is known as a self-
organizing feature map, SOFM.3 Another name commonly used in this context is
Kohonen networks, to honor its inventor.
The Idea Perhaps the best way to explain the nature of SOFM is to use, as a
metaphor, the principle of physical attraction. A code vector, initially generated by
a random-number generator, is subjected to the influence of a sequence of examples
(attribute vectors), each “pulling” the vector in a different direction. In the long run,
the code vector settles in a location that represents a compromise over all these
conflicting forces.
The whole network consists of a set of neurons arranged in a two-dimensional
matrix such as the one shown in Fig. 14.6. Each node (a neuron) in this matrix
represents a code vector that has the same length (the same number of attributes) as
the training examples. At the bottom is an input attribute vector that is connected
to all neurons in parallel. The idea is to achieve by training a situation where
neighboring neurons respond similarly to similar input vectors. For the sake of
simplicity, the input vector in the picture only has two attributes, x1 and x2 . In reality,
it can have a great many.
3
In statistics, and in neural networks, scientists often use the term feature instead of attribute.
14.5 Self-Organizing Feature Maps: Introduction 287
x1 x2
Fig. 14.6 General schema of a Kohonen network
How to Model Attraction Each neuron is described by a weight vector, w D

.w1 ; : : : ; wn / where n is the number of attributes describing the examples. If x D
.x1 : : : ; xn / is the example, and if 2 .0; 1/ is a user-specified learning rate, then
the individual weights are modified according to the following formula:
wi D wi C .xi wi / (14.5)
Note that the i-th weight is increased if xi > wi because the term in the
parentheses is then positive (and is always positive). Conversely, the weight is
decreased if xi < wi because then the term is negative. It is in this sense that we say
that the weight vector is attracted to x. How strongly it is attracted is determined by
the value of the learning rate.
Numeric Example Suppose that an example x D .0:2; 0:8/ has been presented,
and suppose that the winning neuron has the weights w D .0:3; 0:7/. If the learning
rate is D 0:1, then the new weights are calculated as follows:
w1 D w1 C .x1 w1 / D 0:3 C 0:1.0:2 0:3/ D 0:3 0:01 D 0:29
w2 D w2 C .x2 w2 / D 0:7 C 0:1.0:8 0:7/ D 0:7 C 0:01 D 0:71
Note that the first weight originally had a greater value than the first attribute.
By the force of the attribute’s attraction, the weight has been reduced. Conversely,
the second weight was originally smaller than the corresponding attribute, but the
attribute’s “pull” increases it.
Which Weight Vectors Are to Be Attracted by the Example Once an example, x,
has been presented to the neural matrix, a two-step process is launched. The task of
the first step, a “competition,” is to identify in the matrix the neuron whose weight
Fig. 14.7 The idea of “neighborhood” in the Kohonen network
vector is most similar to x. To this end, the Euclidean distance is used—smaller

distance means greater similarity. Once the winner has been established, the second
step updates the weights of this winning neuron as well as the weights of all neurons
in the winner’s physical neighborhood.
A Note on “Neighborhood” Figure 14.7 illustrates what is meant by the neighbor-
hood of the winning code vector, cwinner . Informally, the neighborhood consists of a
set of neurons within a specific physical distance (in the matrix) from cwinner . Note
that this results in a situation where the weights of all neurons in the neighborhood
are modified in a like manner.
Usually, the size of the neighborhood is not fixed. Rather, it is a common
practice to reduce it over time as indicated in the right part of Fig. 14.7. Ultimately,
the neighborhood will degenerate to the single neuron, the one that has won the
competition. The idea is to start with a coarse approximation that is later fine-tuned.
Why Does It Work? The idea motivating the self-organizing feature map is to
make sure that code vectors physically close to each other in the neural matrix
respond to similar examples. This is why the same weight-updating formula is
applied to all neurons in the winner’s neighborhood.
• Describe the general architecture of self-organizing feature maps. Relate the
notion of a code vector to that of a neuron in the matrix.
• Explain the two steps of the self-organizing algorithm: competition, and weight
adjustment. Comment also on the role of the learning rate, .
• What is meant by the neighborhood of the winning code vector? Is its size always
constant?
14.6 Some Important Details 289
14.6 Some Important Details
Having outlined the principle, let us now take a look at some details without which
the technique would underperform.
Normalization The technique does not work well unless all vectors have been
normalized to unit length (i.e., length equal to 1). This applies both to the vectors
describing the examples, and to the weight vectors of the neurons. Fortunately,
normalization to unit length is easy to carry out. Suppose an example is described
as follows:
x D .x1 ; : : : ; xn /
To obtain unit length, we divide each attribute’s value by the length of the original
attribute vector, x:
xi
xi WD q (14.6)
†j xj2
Numeric Example Suppose we want to normalizeq the two-dimensional vector x D

2 2
p p
.5; 5/. The length of this vector is l.x/ D x1 C x2 D 25 C 25 D 50. Dividing
the value of each attribute by this length results in the following normalized version
of the vector:
r r
0 5 5 25 25 1 1
x D .p ; p / D . ; / D .p ; p /
50 50 50 50 2 2
That the length of x0 is qequal to 1 is q

easy to verify: using the Pythagorean
Theorem, we calculate it as x1 C x2 D 12 C 12 D 1. We can see that the new
2 2
attribute vector indeed has unit length.

Initialization The first step in SOFM is the initialization of the neurons’ weights.
Usually, this initialization is carried out by a random-number generator that chooses
the values from an interval that spans equally the positive and negative domains,
say, Œ1; 1. After this, the weight vector is normalized to unit length as explained
above.
Another thing to be decided is the learning rate, . Usually, a small number is
used, say, D 0:1. Sometimes, however, it is practical to start with a relatively high
value such as D 0:9, and then gradually decrease it. The point is to modify the
weights more strongly at the beginning, during the period of early approximation.
After this, smaller values are used so as to fine-tune the results.
The Algorithm The general principle of self-organizing feature maps is summa-
rized by the pseudocode in Table 14.4. To begin with, all examples are normalized
Table 14.4 The basic algorithm of self-organizing feature maps
Input: set of examples without labels

a learning rate, .
a set of randomly initialized neurons arranged in a matrix
1. Normalize all training examples to unit length.

2. Present a training example, and find its nearest code vector, cwinner
3. Modify the weights of cwinner , as well as the weights of the code vectors in the
neighborhood of cwinner , using the formula wi D wi C .xi wi /. After this, re-normalize
the weight vectors.
4. Unless a stopping criterion is met, present another training example, identify cwinner , and
repeat the previous step.
Comments:
1. usually begins with a relatively high value from .0; 1/, then gradually decreases.
2. Every now and then, the size of the neighborhood is reduced.
to unit length. Initial code vectors are created by a random-number generator and
then normalized, too.
In the algorithm’s main body, training examples are presented one by one. After
the presentation of example x, the algorithm identifies a neuron whose weight
vector, cwinner , is the closest to x according to the Euclidean distance. Then, the
weights of cwinner as well as those of all neurons in its neighborhood in the matrix
are modified using Eq. (14.5), and then re-normalized. The algorithm is usually run
for a predefined number of epochs.4
In the course of this procedure, the value of the learning rate is gradually
decreased. Occasionally, the size of the neighborhood is reduced, too.
• Which vectors have to be normalized? What is the goal of this normalization?
Write down the formula that is used to this end.
• What is the subject of initialization? What values are used?
• Summarize the algorithm of self-organizing feature maps.
4
Recall that one epoch means that all training examples have been presented once.
14.7 Why Feature Maps? 291
14.7 Why Feature Maps?
Let us now turn our attention to some practical benefits of self-organizing feature
maps.
Reducing Dimensionality The technique introduced in the previous section essen-
tially maps the original N-dimensional space of the original attribute vectors to
the two-dimensional space of the neural matrix: each example has its winning
neuron, and the winner can be described by its two coordinates in the matrix. These
coordinates can be seen as new description of the original examples.
Creating Higher-Level Features Reducing the number of attributes is of vital
importance in applications where the number of attributes is prohibitively high,
especially if most of these attributes are either irrelevant or redundant. For instance,
this may the case in the field of computer vision where each image (i.e., an example)
can be described by hundreds of thousands of pixels.
Having studied Chap. 7, the reader understands that a direct application of, say,
perceptron learning to attribute vectors of extreme length is unlikely to succeed:
in a domain of this kind, a classifier that does well on a training set tends to
fail miserably on testing data. The mathematical explanation of this failure is that
the countless attributes render the problem’s VC-dimension so high as to prevent
learnability unless the training set is unrealistically large. Advanced machine
learning applications therefore seek to reduce the dimensionality either by attribute
selection or, which is more relevant to this chapter, by way of mapping the multi-
dimensional problem to a smaller space. This is accomplished by creating new
features as functions of the original features.
What New Features Can Thus Be Created One might take this idea a step further.
Here is a simple suggestion, just to offer inspiration. Suppose the training examples
are described by 100 attributes, and suppose we have a reason to suspect that some
attributes are mutually dependent whereas quite a few others are irrelevant.
In such a domain, the dimensionality is likely to be unnecessarily high, and any
attempt to reduce it is welcome. One way to do so is to extract, say, five different
subsets of the attributes, and then redescribe the training examples five times, each
time using a different attribute set. After this, each of these newly obtained training
sets is subjected to SOFM which then maps each multi-dimensional space to two
dimensions. Since this is done five times, we obtain 5 2 D 10 new attributes.
Visualization Human brain has no problem visualizing two- or three-dimensional
data, and then develop an opinion about the similarity (or mutual distance) of
concrete examples, about the examples’ distribution, or about the groups they tend to
form. However, this becomes impossible when there are more than three attributes.
This is when self-organizing feature maps can be practical. By mapping each
example onto a two-dimensional matrix, we may be able to visualize at least some
of the relations inherent in the data. For instance, similar attribute vectors are likely
to be mapped to the same neuron, or at least to neurons that are physically close to
each other. Similar observations can be made about general data distribution. In this
sense, feature maps can be regarded as a useful visualization tool.
Initializing k-Means Each of the weight vectors in the neural matrix can be treated
as a code vector. The mechanism of SOFM makes it possible to find reasonably good
values of these vectors—which can then be used to initialize such cluster-analysis
methods as k-means. Whether this is a practical approach is another question
because SOFM is a computationally expensive technique.
A Brief Mention of “Deep Learning” Let us also remark that methods for
automated creation of higher-level features from very long attribute vectors are
used in so-called deep learning. In essence, deep learning is a neural-networks
technique that organizes the neurons in many layers, many more that we have seen
in the context of multi-layer perceptrons in Chap. 5. It is in this sense the networks
are “deep,” and this is what gave the paradigm its name.
The top layers of these networks are trained by a supervised learning technique
such as the backpropagation of error that the reader already knows. By contrast, the
task for the lower layers is to create a reasonable set of features. It is conceivable
that self-organizing feature maps be used to this end; in reality, more advanced
techniques are usually preferred, but their detailed treatment is outside the scope
of an introductory textbook.
One has to be cautions, though. The circumstance that a concrete paradigm is
popular does not mean that it is a panacea that will solve all machine-learning
problems. Far from it. If the number of features is manageable, and if the size of
the training set is limited, then classical approaches will do just as well as deep
learning, or even better.5
• In what way can the SOFM technique help us visualize the data?
• Explain how the SOFM technique can be used to reduce the number of attributes
and to create new, higher-level features. Where can this be used?
• What is the principle of deep learning?
5
A bulldozer is more powerful than a spade, and yet the gardener prefers the spade most of the
time.
• In some machine-learning tasks, we have to deal with example vectors that do

not have class labels. Still, useful knowledge can be induced for them by the
techniques of unsupervised learning.
• The simplest task in unsupervised learning is cluster analysis. The goal is to find
a way that naturally divides the training set into groups of similar examples. Each
example should be more similar to examples in its group than to examples from
any other group.
• One of the simplest cluster analysis techniques is known under the name of k-
means. Once an initial clustering has been created, the algorithm accepts one
example at a time, and evaluates its distance from the centroid of its own cluster
as well as the distances from the centroids of the other clusters. If the example
appears to find itself in a wrong cluster, it is transferred to a better one. The
technique converges in a finite number of steps.
• The quality of the clusters discovered by k-means is sensitive to initialization and
to the user-specified value of k. Methods to improve this quality by subsequent
merging and splitting, and by alternative initializations sometimes have to be
considered. Also hierarchical implementations of k-means are useful.
• In some domains, the shapes of the data clusters make it impossible for k-means
to find them. For instance, this was the case of the training set shown in Fig. 14.4.
In this event, the engineer will may give preference to some other clustering
technique such as hierarchical aggregation that creates the clusters in a bottom-
up manner, always merging the clusters with the smallest mutual distance.
• In the case of hierarchical aggregation, it is impractical to identify the distance
between two clusters with the distance between their centroids. Instead, we use
the minimum distance between [x,y] where x belongs to one cluster and y to
the other.
• One of the problems facing the k-means algorithm is the question of how to
define the initial code vectors. One way to address this issue is by the technique
of self-organizing feature maps.
• As an added bonus, self-organizing feature maps are capable of converting a
high-dimensional feature space onto only two attributes. They can also be used
for the needs of data visualization.
Historical Remarks The problems of cluster analysis have been studied since the
1960s. The k-means algorithm was described by McQueen [58] and hierarchical
aggregation by Murty and Krishna [71]. The idea of merging and splitting clusters
(not necessarily those obtained by k-means was studied by Ball and Hall [2]. The
technique of SOFM, self-organizing feature maps, was developed by Kohonen [45].
In this book, however, only a very simple version of the technique was presented.
Exercises
1. Look at the three initial clusters of two-dimensional vectors in Table 14.5.

Calculate the coordinates of their centroids.
2. Using the Euclidean distance, decide whether all of the examples from group 1
are where they belong. You will realize that one of them is not. Move it to a more
appropriate group and recalculate the centroids.
3. Normalize the examples in Table 14.5 using Eq. (14.2).
4. Suppose the only information you have is the set of the nine training examples
from Table 14.5, and suppose that you want to run the k-means algorithm for
k D 3 clusters. What will be the composition of the three initial clusters if your
code vectors are .1; 4/; .3; 6/; and .3; 5/?
5. Consider the three clusters from Table 14.5. Which pair of clusters will be merged
by the hierarchical-aggregation technique?
1. At the beginning of this chapter, we specified the benefits of cluster analysis.

Among these was the possibility of identifying neurons in RBF networks with
clusters instead of examples. In the case of k-means, this is straightforward:
each gaussian center is identified with one cluster’s centroid. However, how
would you benefit (in RBF networks) from the clusters obtained by hierarchical
aggregation?
2. Try to invent a machine-learning algorithm that first pre-processes the training
examples using some cluster-analysis technique, and then uses them for classifi-
cation purposes.
Table 14.5 An initial set of Group 1 Group 2 Group 3

three clusters
.1; 4/ .4; 3/ .4; 5/
.3; 6/ .6; 7/ .3; 1/
.3; 5/ .2; 2/ .2; 3/
3. Explain how self-organizing feature maps can be used to define the code vectors
with which k-means sometimes starts. Will it be more meaningful to use the
opposite approach (initialize SOFM) by kmeans)?
1. Write a program that accepts as input a training set of unlabeled examples,

chooses among them k random code vectors, and creates the clusters using the
k-means technique.
2. Write a program that decides whether a pair of clusters (obtained by k-means)
should be merged. The easiest way of doing so is to compare the distance
between the two clusters with the average cluster-to-cluster distance in the given
clustering.
3. Write a program that creates the clusters using the hierarchical aggregation
technique described in Sect. 14.4. Do not forget that the distance between clusters
is evaluated differently than in the case of k-means.
4. Write a program that accepts a set of unlabeled training examples and subjects
them to the technique of self-organizing feature maps.
Chapter 15
Classifiers in the Form of Rulesets
Some classifiers take the form of so-called if-then rules: if the conditions from the
if -part are satisfied, the example is labeled with the class specified in the then-part.
Typically, the classifier is represented not by a single rule, but by a set of rules,
a ruleset. The paradigm has certain advantages. For one thing, the rules capture
the underlying logic, and therefore facilitate explanations of why an example has
to be labeled with the given class; for another, induction of rulesets is capable of
discovering recursive definitions, something that is difficult to accomplish within
other machine-learning paradigms.
In our search for techniques that induce rules or rulesets from data, we will
rely on ideas borrowed from Inductive Logic Programming, a discipline that
studies methods for automated creation and improvement of Prolog programs. Here,
however, we are interested only in classifier induction.
15.1 A Class Described By Rules
To prepare the ground for simple rule-induction algorithms to be presented later,

let us take a look at the nature of the rules we will want to use. After this, we will
introduce some relevant terminology and define the specific machine-learning task.
The Essence of Rules Table 15.1 contains the training set of the “pies” domain
we have encountered earlier. In Chap. 1, the following expression was given as one
possible description of the positive class:
[ (shape=circle) AND (filling-shade=dark) ] OR
[ NOT(shape=circle) AND (crust-shade=dark) ]
When classifying example x, the classifier compares the example’s attribute val-
ues with those in the expression. Thus if x is circular and its filling-shade
happens to be dark, the expression is true, and the classifier therefore labels x with

DOI 10.1007/978-3-319-63913-0_15
298 15 Classifiers in the Form of Rulesets
Table 15.1 Twelve training Crust Filling

examples expressed in a
matrix form
ex3 Triangle Thick Dark Thick Gray pos
ex4 Circle Thin White Thin Dark pos
ex12 Triangle Thick White Thick Gray neg
the positive class. If the expression is false, the classifier labels the example with the
negative class. Importantly, the expression can be converted into the following two
rules:
R1: if [ (shape=circle) AND (filling-shade=dark) ] then pos.
R2: if [ NOT(shape=circle) AND (crust-shade=dark) ] then pos.
else neg.
In the terminology of machine learning, each rule consists of an antecedent (the
if -part), which in this context is a conjunction of attribute values, and a consequent
(the then-part) which points to a concrete class label.
Note that the consequents of both rules indicate the positive class. For an example
to be labeled as positive, it is necessary that the conditions in the antecedent of at
least one rule be satisfied. Otherwise the classifier will label the example with the
default class which, in this case, is neg. We will remember that when working with
rulesets in domains of this kind, one must not forget to specify the default class.
Simplifying Assumptions Throughout this chapter, we will rely on the following
simplifying assumptions:
1. All training examples are described by discrete-valued attributes.
2. The training set is noise-free.
3. The training set is consistent: examples described by the same attribute vectors
must belong to the same class.
The Machine-Learning Task Our goal is an algorithm for the induction of rulesets
from data that satisfy the simplifying assumptions from the previous paragraph. We
will limit ourselves to rules whose consequents point to the positive class, the default
always being the negative class.
Since the training set is supposed to be consistent and noise-free, we will be
interested in classifiers that correctly classify all training examples. This means that
15.1 A Class Described By Rules 299
for each positive example, the antecedent of at least one rule will be true. For any
negative example, no rule’s antecedent is true, and the example is labeled with the
default (negative) class.
A Rule “Covers” An Example Let us introduce one useful term: an example either
is or is not covered by a rule. A simple illustration will clarify the notion. Consider
the following rule:
R: if (shape=circle) then pos.
If we apply this rule to the examples from Table 15.1, we will observe that
the antecedent’s condition, shape=circle, is satisfied by the following set
of examples: fex1 ; ex2 ; ex4 ; ex6 ; ex7 ; ex0 g. We will say that R covers these six
examples. Generally speaking, a rule covers an example if the expression in the
rule’s antecedent is true for this example. Note that four of the examples covered by
this particular rule are positive and two are negative.
Rule Specialization Suppose we modify the above rule by adding to its antecedent
another condition, filling-shade=dark, obtaining the following:
R1: if (shape=circle) AND (filling-shade=dark) then pos
Checking R1 against the training set, we realize that it covers the following
examples: fex1 ; ex2 ; ex4 ; ex6 g. We observe that this is a subset of the six examples
originally covered by R. Conveniently, only positive (and no negative) examples are
now covered.
This leads us to the definition of another useful term. If a modification of a
rule’s antecedent reduces the set of covered examples to a subset, we say that the
modification has specialized the rule. In other words, specialization narrows the set
of covered examples to a proper subset. A typical way of specializing a rule is to
add a new condition to the rule’s antecedent.
Rule Generalization Conversely, a rule is generalized if its modification
enlarges the set of covered examples to a superset—if the new version covers
all examples that were covered by the previous version, plus some additional
ones. The easiest way to generalize a rule is by removing a condition from its
antecedent. For instance, this happens when we drop from rule R1 the condition
(filling-shade=dark).
Specialization and Generalization of Rulesets We have said we are interested in
induction of rulesets that label an example with the positive class if the antecedent
of at least one rule is true for the example. For instance, this is the case of the ruleset
consisting of the rules R1 and R2 above.
If we remove one rule from a ruleset, the ruleset may no longer cover some of
the previously covered examples. This, we already know, is called specialization.
Conversely, adding a new rule to the ruleset will generalize the ruleset because the
new rule will add to the set of covered examples.
• Explain the nature of rule-based classifiers. What do we mean when we say
that a rule covers an example? Using this term (cover), specify how the induced
classifier should behave on a consistent and noise-free training set.
• Define the terms generalization and specialization. How will you specialize or
generalize a rule? How will you specialize or generalize a ruleset?
• List the simplifying assumptions to be used throughout this chapter.
15.2 Inducing Rulesets by Sequential Covering
Let us now introduce a simple technique that induces rulesets from training data
satisfying the simplifying assumptions from the previous section.
The Principle The goal is to find a ruleset such that each of its rules covers some
positive examples, but no negative examples. Together, the rules should cover all
positive examples and no negative ones. The procedure we will use creates one rule
at a time, always starting with a very general initial version (covering also negative
examples) that is then gradually specialized until all negative examples are excluded
from coverage. The circumstance that the rules are created sequentially, and that
each is supposed to cover those positive examples that were missed by previous
rules, gives the technique its name: sequential covering.
Baseline Version of Sequential Covering Table 15.2 provides the pseudocode of
a simple method for induction of rulesets. The main body contains the sequential
covering algorithm. The idea is to find a rule that covers some positive examples,
but no negative examples. Once the rule has been created, the examples it covers are
removed from the training set. If no positive examples remain, the algorithm stops;
otherwise, the algorithm is applied to the reduced training set.
The lower part describes induction of a single rule. The algorithm starts with
the most general version of the antecedent that says, “all examples are positive.”
Assuming that the training set contains at least one negative example, this statement
is obviously incorrect. The algorithm therefore seeks to rectify the situation by
specialization, trying to exclude from coverage some negative examples, hopefully
without losing the coverage of the positive examples. The specialization operator
adds to the rule another conjunct in the form, ai D vj (read: the value of attribute ai
is vj ).
A Concrete Example Let us “hand-simulate” the sequential-covering algorithm
using the data from Table 15.1. The first rule, with the empty antecedent, covers all
training examples. Adding to the empty antecedent the condition shape=circle
15.2 Inducing Rulesets by Sequential Covering 301
Table 15.2 The sequential covering algorithm

Input: training set T.
Sequential covering.
Create an empty ruleset.
While at least one positive example remains in T:
1. Create a rule using the algorithm below.

2. Remove from T all examples that satisfy the rule’s antecedent.
3. Add the rule to the ruleset.
Create a single rule

Create an initial version of the rule, R: if () then pos
1. If R does not cover any negative example, stop.

2. Add to R’s antecedent a condition, ai D vj , and return to the previous step.
results in a rule that covers four positive and two negative examples. Adding one
more condition, filling-shade=dark, specializes the rule so that, while still
covering the four positive examples, it now no longer covers any negative example.
We have obtained a rule that covers examples fex1 ; ex2 ; ex4 ; ex6 g. Note that is the
rule R1 from the previous section.
If we remove these four examples from the training set, we are left with only
two positive examples, ex3 and ex5 . The development of another rule again starts
from the most general version (empty antecedent). Suppose that we then choose
shape=triangle as the initial condition. This covers one positive and two
negative examples. Adding to the antecedent the term filling-shade=dark,
we succeed in excluding the negative examples while retaining the coverage of the
positive example ex3 , which can now be removed from the training set. After the
creation of this second rule, we are left with one positive example ex5 .
We therefore have to create yet another rule whose task will be to cover ex5
without covering any negative example. Once we find such rule, ex5 is removed
from the training set. After this, we observe that there are no positive examples left,
and the procedure can stop. We have created a ruleset consisting of three rules that
cover all positive examples and no negative examples.
How to Identify the Best Attribute-Value Pair In the previous example, we
always chose the condition to be added to the rule’s antecedent more or less at
random. But seeing that we could have selected it from quite a few alternatives, we
realize that we need a mechanism capable of informing us about the quality of each
choice. Perhaps the most natural criterion to be used here is based on information
theory, a principle we have encountered in Chap. 6 where we used it in the course
of induction of decision trees.
C
Let Nold be the number of positive examples covered by the original version of

the rule, and let Nold be number of negative examples covered by the original version
of the rule. Likewise, the numbers of positive and negative examples covered by the
C
new version of the rule will be denoted by Nnew and Nnew , respectively.
Since the rule covers only positive examples, the information content of the
message that a randomly picked example is labeled by it as positive is calculated
as follows (for the old version and for the new version):
C
Nold
Iold D log. C
/
Nold C Nold
C
Nnew
Inew D log. C C N
/
Nnew new
The difference between these two is the amount of information that has been
gained by modifying the rule. Usually, machine-learning professionals normalize
the information gain by the number, NC , of covered examples so as to give
preference rule modifications that optimize the number of covered examples.
The quality of the rule-improvement is then calculated as follows:
Q D NC jInew Iold j (15.1)
When comparing alternative ways of modifying a rule, we choose the one with
the highest value of Q.
• Summarize the principle of the sequential covering algorithm.
• Explain the mechanism of gradual specialization of a rule. What do we want to
accomplish by this specialization?
• How will you use information gain when looking for the most promising way of
specializing a rule?
15.3 Predicates and Recursion
The sequential covering algorithm has a much broader scope of applications than the
previous section seems to have indicated. Perhaps most importantly, the technique
can be employed for induction of concepts expressed in predicate calculus.
15.3 Predicates and Recursion 303
Predicates: Greater Expressive Power Than Attributes A serious limitation of

attribute-value logic is that it is not sufficiently flexible to capture certain relations
among data. For instance, the fact that y is located between x and z can be stated
using the predicate between(x,y,z)—more accurately, the predicate is the term
“between,” whereas “(x,y,z)” is a list of the predicate’s arguments.
The reader will agree that trying to express the same relation by means of
attributes and their values would be difficult to say the least. An attribute can be
seen as a special case of a one-argument predicate. For instance, the fact that, for a
given example, x, the shape is circular can be written as circular(x). But
the analogy is no longer as obvious in the case of predicates with more arguments.
Induction of Rules in Predicate Calculus Here is an example of a rule that says
that if x is a parent of y, and at the same time x is a woman, then this parent is
actually y’s mother:
if parent(x,y) AND female(x) then mother(x,y)
We can see that this rule has the same structure as the rules R1 and R2 we
have seen above: a list of conditions in the antecedent followed by a consequent.
And indeed, the same sequential covering algorithm can be used here. There is
one difference, though. When choosing among candidate predicates to be added
to antecedent, we must not forget that the meaning of the predicate changes if we
change the arguments. For instance, the previous rule’s meaning will change if we
replace parent(x,y) with parent(x,z) because, in this case, the fact that x
is a parent of z surely does not guarantee that x is mother of some other subject, y.
Rulesets Allow Recursive Definitions The rules can be more interesting than
the toy domain from Table 15.1 might lead us to believe. For one thing, they
can be recursive—which is the case of the following two rules defining the term
ancestor.
if parent(x,y) then ancestor(x,y).
if parent(x,z) AND ancestor(z,y) then ancestor(x,y).
The meaning of two rules is easy to see. Ancestor is a parent, or at least
the parent’s ancestor. For instance, a grandparent is the parent of a parent—and
therefore an ancestor.
A Concrete Example of Induction Let us illustrate induction of rulesets using the
problem from Table 15.3. Here, two concepts (classes), parent and ancestor,
are characterized by a list of positive examples under the assumption that any
example that is not in this list should be regarded as a negative example. Our goal is
to induce the definition of ancestor, using the predicate parent.
We begin with the most-general rule, if () then ancestor(x,y). In the next
step, we want to add a condition to the antecedent. To this end, we may consider
various possibilities, but the simplest appears to be parent(x,y)—which will
also be supported by the information-gain criterion. We have obtained the following
rule:
Table 15.3 Illustration of induction from examples described using predicate logic
Consider the knowledge base consisting of the following positive examples of classes
parent and ancestor, defined using prolog-like facts (any other example will be
regarded as negative).
parent(eve,ted) ancestor(eve,ted)ancestor(eve,ivy)
parent(tom,ted) ancestor(tom,ted)ancestor(eve,ann)
parent(tom,liz) ancestor(tom,ted)ancestor(eve,jim)
parent(ted,ivy) ancestor(tom,ted)ancestor(tim,ivy)
parent(ted,ann) ancestor(tom,ted)ancestor(eve,ann)
parent(ann,jim) ancestor(tom,ted)ancestor(eve,jim)
ancestor(ted,jim)
From the above examples, the algorithm creates the following first version of the rule. Note
that this rule does not cover any negative examples.
R3: if parent(x,y) then ancestor(x,y)
Removing all positive examples of this rule, the following set of positive examples of
ancestor(x,y) remains:
ancestor(eve,ivy)
ancestor(eve,ann)
ancestor(eve,jim)
ancestor(tim,ivy)
ancestor(eve,ann)
ancestor(eve,jim)
To cover these, another rule is created:
if parent(x,z) then ancestor(x,y)
After specialization, the second rule is turned into following:
R4: if parent(x,z) AND ancestor(z,y) then ancestor(x,y)
These two rules R3 and R4 now cover all positive examples and no negative examples.
R3: if parent(x,y) then ancestor(x,y)

Observing that the rule covers only positive examples and no negative examples,
we realize there is no need to specialize it.
However, the rule covers only the ancestor examples from the middle column,
and no examples from the rightmost column. Obviously, we need at least one more
rule. When considering the conditions to be added to the empty antecedent of the
15.4 More Advanced Search Operators 305
next rule, we may consider the following (note that this is always the same predicate,
but each time with a different set of arguments):
parent(x,z)
parent(z,y)
Suppose that the first leads to higher information gain. Seeing that the rule
still covers some negative examples, we want to specialize it by adding another
condition to its antecedent. Seeing that the parent predicate does not lead us
anywhere, we try the predicate ancestor, again with various lists of arguments.
Evaluating the information gain of all alternatives, we realize that the best option is
ancestor(z,y). This is how we obtain the second rule:
R4: if parent(x,z) AND ancestor(z,y) then ancestor(x,y).
• How can a class be expressed using predicates? In what sense is the language of
predicates richer than the language of attributes?
• Give an example of a recursively defined class. Can you think of a different
example than ancestor?
15.4 More Advanced Search Operators
The technique described in the previous sections followed a simple strategy: in its
attempts to find a good ruleset, the algorithm always sought to modify the rule(s) by
specialization and generalization, evaluating alternative options by the information-
gain criterion.
Operators for Ruleset Modification In reality, the search for rules can be more
flexible than that. Other ruleset-modifying operators have been suggested. These,
as we will see, do not necessarily represent specialization or generalization, but if
we take a look at them, we realize they make sense. Let us mention in passing that
these operators have been derived with the help of a well-known principle from
logic, so-called inverse resolution. For our specific needs, however, the method of
their derivation is unimportant.
In the following, we will simplify the formalism by writing a comma instead of
AND, and using an arrow instead of the if-then construct. In all of the four cases,
the operator converts the ruleset on the left into the ruleset on the right. The leftmost
column gives the traditional names of these operators.

b; x ! a b; x ! a
identification: )
b; c; d ! a c; d ! x

c; d ! x c; d ! x
absorption: )
b; c; d ! a b; x ! a
8 9
< u; b; c ! a =
v; b; c ! a
inter-construction: ) v!u
w; b; c ! a : ;
w!u
8 9
< v; u ! a =
v; b; c ! a
intra-construction: ) w; u ! a
w; b; c ! a : ;
b; c ! u
Note that these replacements are not deductive: the rules on the right are never
perfectly equivalent to those on the left. And yet, they do appear to make sense
intuitively.
How to Improve Existing Rulesets? The operators from the previous paragraph
can be used to improve rulesets that have been induced by the sequential covering
algorithm. We can even consider a situation where not one, but several different
classes were induced, which gave rise to several rulesets.
These rulesets can then be improved applying the hill-climbing search technique.
The search operators are those listed in the previous paragraph. The evaluation
function may give preference to more compact rules that classify correctly some
auxiliary set of training examples meant to represent a concrete application domain.
• List the ruleset-modifying operators listing in this section. Which field of logic
has helped derive them?
• Suggest how you might use these rules in an attempt to improve a given ruleset?
• Some classifiers have the form of rules. A rule consists of an antecedent (a list of
conjuncted conditions) and a consequent (a class label). If the rule’s antecedent
is true, for the given example, then the example is labeled with the label pointed
to by the consequent.
• If a rule’s antecedent is true, for an example, we say that the rule covers the
example.
• In the course of rule induction, we often rely on specialization. This reduces the
set of covered examples to its subset. A rule is specialized if we add a condition
to its antecedent. Conversely, generalization enlarges the set of covered examples
to its superset.
• Usually, we induce a set of rules, a ruleset. The classifier then labels an example
as positive if the antecedent of at least one of the rules is true. Adding a
rule to a ruleset represents generalization. Removing a rule would represent
specialization.
• The chapter introduced a simple algorithm for induction of rulesets from noise-
free and consistent training data described by discrete attributes. The algorithm
can so to some degree be optimized with the help of a criterion derived from
information theory.
• The same algorithm can be used for induction of rules in domains where the
examples are described using predicate calculus. Even recursive rules can thus
be discovered.
• Some other “search operators” have been developed by the field of inverse
resolution. They do not necessarily represent specialization or deduction.
Historical Remarks Induction of rules belongs to the oldest tasks of machine
learning since the days when this discipline was seen as a means of inducing
knowledge artificial-intelligence systems. The sequential-covering algorithm is a
simplified version of an algorithm by Clark and Niblett [16]. Its use for induction of
predicate-based rule was inspired by the FOIL algorithm developed by Quinlan [77].
The additional operators from Sect. 15.4 are based on the operators introduced by
Muggleton and Buntine [70] in the framework of their work on inverse resolution.
Exercises
1. Hand-simulate the algorithm of sequential covering for the data from Table 15.1.
Ignoring information gain, indicate how the first rule is created if we start from
crust-shade=gray.
2. Show that, when we choose different ways of specializing a rule (adding different
attribute-value pairs), we in the end obtain a different ruleset, often of a different
size.
1. Think of some other examples of classes (different from those discussed in this
chapter) that are best defined recursively.
2. Think about how classes that are by nature recursive would be difficult to address
in the framework of attribute-value logic. Demonstrate the superior power of the
predicate calculus.
3. Suggest a learning procedure for “knowledge refinement.” In this task, we
assume that certain classes have already been defined in predicate calculus. When
presented with another set of examples, the knowledge-refinement technique
seeks to optimize the existing rules, either my making them more compact, or
by making them more accurate in the presence of noise.
1. Write a computer program that implements the sequential covering algorithm.

Use some simple criterion (not necessarily information gain) to choose which
condition to add to a rule’s antecedent.
2. In the UCI repository, find a domain satisfying the criteria specified in Sect. 15.1.
Apply to it the program developed in the previous step.
3. How would you represent two-argument or three-argument predicates if you
wanted to implement your machine-learning program in C++, Java? or some
other programming language of a similar nature?
4. Write a program that applies the sequential covering algorithm to examples
described in predicate calculus.
Chapter 16
The Genetic Algorithm
The essence of machine learning is the search for the best solution to our problem:
to find a classifier which classifies as correctly as possible not only the training
examples, but also future examples. Chapter 1 explained the principle of one of the
most popular AI-based search techniques, the so-called hill-climbing, and showed
how it can be used in classifier induction.
There is another approach to search: the Genetic Algorithm, inspired by the
principles of Darwinian evolution. The reader needs to be acquainted with it because
the technique can be very useful in dealing with various machine-learning problems.
This chapter presents the baseline version, and then illustrates its use using certain
typical issues from the field of k-NN classifiers.
16.1 The Baseline Genetic Algorithm
Let us first briefly describe the general principle of the genetic algorithm, relegating
the details of implementation to the next section.
The Basic Philosophy In this section, the classifier will encode in the form of a
chromosome, which most of the time will be a string of bits that are sometimes
referred to as “genes.” The genetic algorithm operates with a population of
chromosomes, each describing one individual (a classifier). Each such individual
is assigned a value by a fitness function; this value will usually depend on the
classifier’s performance. The fitness function plays a role analogous to that of the
evaluation function in heuristic search.1
1
This chapter will use the terms “evaluation function,” “survival function”, and “fitness function”
interchangeably.

DOI 10.1007/978-3-319-63913-0_16
310 16 The Genetic Algorithm
Fig. 16.1 The genetic

algorithm’s endless loop.
Each individual in the
population has its chance of
survival. Recombination of
the genetic information mutation
provided by mating partners a
creates new chromosomes
population recombination
that may be corrupted by
mutation
wheel of mating
fortune
survivors
The Genetic Algorithm’s Loop The genetic algorithm operates in an endless loop
depicted in Fig. 16.1. At each moment, there is a population of individuals, each
with a certain value of the fitness function. This value then determines the size of
the segment belonging to the individual in a “wheel of fortune” that determines the
individual’s chances of survival. It is important to understand the probabilistic nature
of the process. While an individual with a larger segment enjoys a higher chance of
survival, there is no guarantee of it because the survival game is non-deterministic.
In the real world, too, a specimen with excellent genes may perish in a silly accident,
while a weakling can make it by mere good luck. But in the long run, and in large
populations, the laws of probability will favor genes that contribute to high fitness.
The surviving specimens will then choose “mating partners.” In the process
of mating, the chromosomes of the participating individuals are recombined (see
below), which gives rise to a pair of new chromosomes. These new chromosomes
may subsequently be subjected to mutation, which essentially adds noise to the
strings of genes.
The whole principle is summarized by the pseudocode in Table 16.1.
How the Endless Loop Works Once a new population has been created, the
process enters a new cycle in which the individuals are subjected to the same
wheel of fortune, followed by mating, recombination, and mutation, and the story
goes on and on until stopped by an appropriate termination criterion. Note how
occasional wrong turns are eliminated by the probabilistic nature of process.
A low-quality chromosome may survive the wheel of fortune by a fluke; but if its
children’s fitness values remain low, the genes will perish in subsequent generations
16.2 Implementing the Individual Modules 311
Table 16.1 The principle of the genetic algorithm

initial state: a population of individual chromosomes
1. The fitness of each individual is evaluated. Based on its value, individuals are
randomly selected for survival.
2. Survivors select mating partners.
3. New individuals are created by chromosome recombination of the mating partners.
4. Individual chromosomes are corrupted by random mutation.
5. Unless a termination criterion is satisfied, the algorithm returns to step 1.
anyway. Alternatively, some of an unpromising individual’s genes may prove to be

useful when embedded in different chromosomes which they may enter through
recombination. By giving them an occasional second chance, the process offers
flexibility that would be impossible in a more deterministic setting.
• Explain the main principle of the genetic algorithm. How are the individuals
described here? What is meant by their “survival chances”?
• Summarize the basic loop of the genetic algorithm.
• What is the advantage of the probabilistic implementation of the principle of
survival as compared to a possible deterministic implementation?
16.2 Implementing the Individual Modules
Let us take a closer look at how to implement in a computer program the basic
aspects of the genetic algorithm: the survival game, the mating process (partner
selection), chromosome recombination, and mutation. To begin with, we will
discuss only very simple solutions, relegating more advanced techniques to later
sections.
For the sake of simplicity, we will assume that the chromosomes acquire the
form of binary strings such as [1 1 0 1 1 0 0 1], where each bit represents
a certain property that is either present, in which case the bit has value 1, or absent,
in which case the bit is 0. Thus in a simplified version of the “pies” problem, the first
bit may indicate whether or not the crust is thick, the second bit may indicate
whether or not the filling is black, and so on.
Initial Population The most common approach to creating the initial population
will employ a random-number generator. Sometimes, the engineer can rely on some
knowledge that may help her create initial chromosomes known to outperform
randomly generated individuals. In the “pies” domain, this role can be played by
the descriptions of the positive examples. However, one has to make sure that the
initial population is sufficiently large and has sufficient diversity.
The Survival Game The genetic algorithm assumes that there is a way to calculate
for each specimen its survival chances. In some applications, these chances can be
established by a practical experiment that lets the individual specimens to fight it out.
In other domains, the fitness is calculated by a user-specified evaluation function
whose value depends on the chromosome’s properties. And if the chromosome
represents a classifier, the fitness function can rely on the percentage of the training
examples correctly labeled by the classifier.
An individual’s survival is determined probabilistically. Here is how to imple-
ment this “wheel of fortune” in a computer program. Let Fi denote the i-th
specimen’s fitness and let F D †i Fi be the sum of all individual’s fitness values
that are then arranged along the interval .0; F. The survival is modeled by a
random-number generator that returns some r 2 .0; F: the sequential number of
the subinterval that has been “hit” by r then points to the survivor. The principle
is illustrated in Fig. 16.2 for a small population of four specimens and a random
number that lands in the third interval so that individual 3 is selected. If the
fate wants 20 specimens to survive, it has to generate 20 random numbers whose
locations in the interval .0; F identify the survivors.
Whereas specimens with small fitness are likely to get eliminated, those with
higher values can appear in the pool of survivors more than once. A biologist will
wince at this “cloning” idea, but in the pragmatic world of computer programmers,
the same individual can “survive” twice, three times, or even many times.
The Mating Operator The survival game is followed by mating. In nature, an
individual judges a partner’s suitability by strength, speed, or sharp teeth. Something
similar is accomplished in a computer implementation by means of the fitness
random number
0 8 13 22 25
individual 1 individual 2 individual 3 individual 4
Fig. 16.2 The axis represents a population of four individuals whose fitness values are 8; 5; 9; and
3; respectively. Since the randomly generated number, 15, falls into the third subinterval, the third
individual is selected
16.2 Implementing the Individual Modules 313
function. There is a difference, though: the notion of sex is usually ignored—any

chromosome can mate with any other chromosome.
An almost trivial mating strategy will pair the individuals arbitrarily, perhaps
generating random pairs of integers from the interval Œ1; Ns , where Ns is the number
of specimens in the population. However, this technique fails to do justice to
the circumstance that specimens with high fitness are likely to be deemed more
attractive than others. A simple way to reflect this in a computer program is to order
the individuals in a descending order of their fitnesses, and then pair the neighbors.
Yet another strategy does it probabilistically. It takes the highest-ranking indi-
vidual, then chooses its partner using the mechanism employed in the survival
game—see Fig. 16.2. The same is done for the second highest-ranking individual,
then for the third, and so on, until the new population has reached the required
size. “Better” individuals are thus likely (though not guaranteed) to mate with
other strong individuals. Sometimes, the partner will have low value (due to the
probabilistic selection), but this gives rise to diversity that gives the system the
opportunity to preserve valuable chromosome chunks that only have the bad luck
of being currently incorporated in low-quality specimens.
Long-Living and Immortal Individuals One of the shortcomings of this algo-
rithm is that a very good organism may be replaced by lower-valued children
and useful genes may disappear. To prevent this from happening, some computer
programs copy the best specimens into the new generation alongside their children.
For instance, the program may directly insert in the new generation 20% best
survivors, and then create the remaining 80% by applying the recombination and
mutation operators to the best 95% individuals, totally ignoring the bottom 5%. In
this way, not only will the best specimens live longer (even become “immortal”),
but the program will also get rid of some very weak specimens that have survived
by mere chance.
Chromosome Recombination: One-Point Crossover The simplest way to imple-
ment chromosome recombination is by the one-point crossover, an operator that
swaps parts of the information in the parent chromosomes. The principle is simple.
Suppose that each chromosome consists of a string of n bits and that a random-
number generator has returned an integer i 2 Œ1; n. Then, the last i bits in the first
chromosome (its i-bit tail) are replaced with the last i bits in the second chromosome
and vice versa. A concrete implementation can permit the situation where i D n, in
which case the two children are just replications of their parents. In the example
below, the random integer is i D 4, which means that 4-bit tails are exchanged (the
crossover point is indicated by a space).
1101 1001 1101 0111
)
0010 0111 0010 1001
The reader can see that the children tend to take after their parents, especially
when the exchanged tails are short. The maximum distance between the children
and the parents is achieved when i D n 1.
In many applications, the recombination operator is applied only to a certain

percentage of individuals. For instance, if 50 pairs have been selected for mating,
and if the probability of recombination has been set by the user as 80%, then only
40 pairs will be subject to recombination, and the remaining 10 will just be copied
into the next generation.
The Mutation Operator The task for mutation is to corrupt the inherited genetic
information. Practically speaking, this is done by flip-flopping a small percentage of
the bits in the sense that a bit’s value 0 is changed to 1 or the other way round. The
concrete percentage (the frequency of mutations) is a user-set parameter. Suppose
that this parameter requires that p D 0:001 of the bits should on average be thus
affected. The corresponding program module will then for each bit generate a
random integer from the interval Œ1; 1000. If the integer equals 1, then the bit’s
value is changed, otherwise it is left alone.
Let us give some thought to what frequency of mutations we need. At one
extreme, very rare mutations will hardly have any effect at all. At the other extreme,
very high mutation frequency would disrupt the genetic search by damaging too
many chromosomes. If the frequency approaches 50%, then each new chromosome
will behave as a randomly generated bit string; the genetic algorithm then degener-
ates to a random-number generator.
The mutation operator serves a different purpose than the crossover operator. In
the one-point crossover, no new information is created, only existing substrings are
swapped. Mutation introduces some new twist, previously absent in the population.
• What is the main task of the survival game and how would you implement it in a
computer program?
• Describe a simple mechanism to implement the selection of the mating partners.
Describe the recombination operator, and the mutation operator.
16.3 Why It Works
Let us now offer an intuitive explanation of the genetic algorithm’s performance.

Function Maximization The goal of the simple problem in Table 16.2 is to find
the value of x for which the function f .x/ D x2 x is maximized. Each chromosome
in the second column of the upper table is interpreted as a binary-encoded integer
whose decadic value is given in the third column. The fourth column gives the
16.3 Why It Works 315
Table 16.2 Illustration of the genetic algorithm

Suppose we want the genetic algorithm to find the maximum of f .x/ D x2 x. Let x be
an integer represented by a binary string. The initial population consists of the four
strings in the following table that for each of them gives the integer value, x, the
corresponding f .x/, the survival chances (proportional to f .x/), and the number of times
each exemplar was selected for the next generation.
Initial x x2 x Survival actual

No. population chance count
1 01100 12 132 0.14 1
2 11001 25 600 0.50 2
3 01000 8 56 0.05 0
4 10011 19 342 0.31 1
Average 282
Maximum 600
In the sample run reported here, the neighboring specimens mated, exchanging 1-bit tails
and 3-bit tails, respectively, as dictated by the randomly generated tail lengths (the
crossover sites indicated by spaces). No mutation is used here. The last two columns give
the values of x and f .x/ for the new generation.
After Mate Tail New x x2 x

reproduction with length population
0110 0 2 1 01101 13 156
1100 1 1 1 11000 24 552
11 001 4 3 11011 27 702
10 011 3 3 10001 17 289
Average 425
Maximum 702
The reader can see that the value of the best specimen and the average value in the entire
population have increased.
corresponding f .x/ whose relative value, shown in the fifth column, then determines
for each individual its survival chances. For example, the first specimen has f .x/ D
122 12 D 132 and the relative chances of survival (in this particular population)
are 14% because 132=.132 C 600 C 56 C 342/ D 0:14. The rightmost column
tells us how many times each individual has been selected for inclusion in the next
generation.
In the next step, the survivors identify their mating partners. Let us assume that
we have simply paired the neighboring specimens: the first with the second, and the
third with the fourth. Then, the random selection of the crossover point dictates that
1-bit tails be exchanged in the first pair and 3-bit tails in the second. No mutation
is applied. The result is shown in the bottom table where the last three columns
Fig. 16.3 After exchanging f(x)

4-bit tails, two parent
chromosomes (upper strings)
give rise to two children
(lower strings). There is a
chance that at least one child
will “outperform” both
parents
0011 1110 0101 1001 x

0011 1001 0101 1110
show, respectively, the new binary strings, their decadic values, and the values of
f .x/. Note that both the average and the maximum value of the fitness function have
increased.
Do Children Have to Outperform Their Parents? Let us ask what caused this
improvement. An intuitive answer is illustrated in Fig. 16.3 that shows the location
of two parents and the values of the survival function, f .x/, for each of them
(the dashed vertical lines). When the two chromosomes swap their 4-bit tails, two
children are created, each relatively close to one of the parents. The fact that each
child finds itself in a region where the values of f .x/ are higher than those of the
parents begs the question: are children always more fit than their parents? Far from
that. All depends on the length of the exchanged tails and on the shape of the fitness
function. Imagine that in the next generation the same two children get paired with
each other and that the randomly generated crossover point is at the same location.
Then, these children’s children will be identical to the two original strings (their
“grandparents”); this means that the survival chances decreased back to the original
values. Sometimes, both children outperform their parents; in other cases, they are
weaker than their parents; and quite often, we get a mixed bag. What matters is that
in a sufficiently large population, most of the better specimens will survive because
the selection process favors individuals with higher fitness, f .x/. Unfit specimens
will occasionally make it, but they tend to lose in the long run.
If the exchanged string-tails are short, the children are close to their parent
chromosomes. Long tails will give rise to children much less similar to their parents.
As for mutation, its impact on the distance between the child and its parent depends
on which bit is mutated. If it is the leftmost bit, the mutation will cause a big jump
along the horizontal axis. If it is the rightmost bit, the jump is short. Either way,
mutation complements recombination. Whereas the latter tends to explore the space
in the vicinity of the parent chromosomes, the former may look elsewhere.
The Shape of the Fitness Function Some potential pitfalls inherent in the
definition of the fitness functions are illustrated in Fig. 16.4. The function on the
left is almost flat. The fact that different individuals have here virtually the same
chances to survive defeats the purpose of the survival game. When the survivors are
16.4 The Danger of Premature Degeneration 317
f(x) f(x)
x x
Fig. 16.4 Examples of two fitness functions that are poor guides for the genetic search. To be
useful, the survival function should not be too flat and it should not contain isolated narrow peaks
chosen according to a near-uniform distribution, the qualities of the individuals will

not give these individuals any perceptible competitive advantage. This drawback
can be mitigated by making f .x/ less flat. There is an infinite number of ways this
can be achieved, one possibility being to replace f .x/ with, say, f .x/ D f 2 .x/.
The right-hand part of Fig. 16.4 shows another pitfall: isolated narrow peaks. In
comparison to the widths of the “peaks,” children may find themselves too far from
their parents. For instance, if the parent lies just at a hill’s foot, the child may find
itself on the opposite side, in which case the peak will go unnoticed. This problem
is more difficult to prevent than the previous one.
• Explain how the location of the crossover point determines how much the
children will differ from their parents.
• Explain how the mutual interplay between recombination and mutation may
affect the survival chances. Show how they also depend on the concrete shape
of the survival function and on the location of the parents.
16.4 The Danger of Premature Degeneration
The fact that the genetic algorithm reached a value that does not seem to improve
over a series of generation does not yet mean the search has been successful. The
plateau may be explained by other circumstances.
Premature Degeneration A simple implementation of the genetic algorithm will
stop after a predefined number of generations. A more sophisticated version will
keep track of the highest fitness value achieved so far, and then terminate the search
when this value no longer improves.
There is a catch, though. The fact that the fitness value has reached a plateau may
not guarantee that a solution has been found. Rather, the search might have reached
the stage called premature degeneration. Suppose that the search from Table 16.2
has reached the following population:
0 1 0 0 0
0 1 0 0 1
0 1 0 0 0
0 1 0 0 0
What are the chances of improving this population? Recombination will not get
us anywhere. If the (identical) last two chromosomes mate, the children will only
be copies of the parents. If the first two are paired, then 1-point crossover will only
swap the rightmost bit, an operation that does not create a new chromosome, either.
The only way to cause a change is to use mutation. By changing the appropriate bits,
mutation can reignite the search. For instance, this will happen after the mutation
of the third bit in the first chromosome and the fourth bit (from the left) of the last
chromosome. Unfortunately, mutations are rare, and to wait for this to happen may
be impractical. For all practical purposes, premature degeneration means the search
got stuck.
Preventing Premature Degeneration Premature degeneration has a lot to do with
the population’s diversity. The worst population is one in which all chromosomes
have exactly the same bit string, something the engineer wants to avoid. Any
computer implementation will therefore benefit from a module that monitors
diversity and takes action whenever it drops below a certain level. A simple way
to identify this situation is to calculate the average similarity between pairs of
chromosomes, perhaps by counting the number of bits that have the same value
in both strings. For instance, the similarity between [0 0 1 0 0] and 0 1 1 0
0] will be 4 (four bits are equal) and the similarity between [0 1 0 1 0] and
[1 0 1 0 1] will be 0.
Once a drop in average chromosome-to-chromosome similarity has been
detected, the system has to react. This is not yet a cause for alarm. Thus in
the function-maximization example, advanced generations will be marked by
populations where most specimens are already close to the maximum. This kind of
“degeneration” will certainly not be deemed “premature.” However, the situation
is different if the best chromosome can be shown to be very different from the
solution. In this event, we have to increase diversity.
Increasing Diversity Several strategies can be used. The simplest will just insert
in the current population one or more newly created random individuals. A more
sophisticated approach will run the genetic algorithm on two or more populations in
parallel, in isolation from each other. Then, either at random intervals, or whenever
16.5 Other Genetic Operators 319
premature degeneration is suspected, a specimen from one population will be

permitted to choose its mating partner in a different population. When implementing
this technique, the programmer has to decide in which population to place the
children.
The Impact of Population Size Special attention has to be paid to the size of the
population. Usually, though not always, the size is kept constant throughout the
entire genetic search. The number of individuals in the population will be dictated
by the concrete application. As a rule of thumb, smaller populations will need
many generations to reach a good solution—unless they degenerated prematurely.
Very large populations may be robust against degeneration, but they may incur
impractical computational costs.
• In what way does the success of the genetic algorithm depend on the definition
of the fitness function? What are the two main pitfalls? How would you handle
them?
• What criteria to terminate the genetic search would you recommend? What are
their advantages and disadvantages?
• What is premature degeneration? How can it be detected and how can the
situation be rectified? Why do we need diversity in the population?
• Discuss the impact of the population size.
16.5 Other Genetic Operators
We have introduced only a very simple version of the genetic algorithm and its
operators. Now that the reader understands the principle, we take a look at some
alternatives.
Two-Point Crossover The one-point crossover introduced above is only a special
case of the much more common two-point crossover. Here, the random-number
generator is asked to return two integers that define two locations in the binary
strings. The parents then swap the substrings between these two locations as
illustrated below (the two crossover points are indicated by spaces).
110 110 01 110 001 01
)
001 001 11 001 110 11
The two crossover points can be different for each chromosome. In this event,
each parent will “trade” a different substring of its chromosome as indicated below.
1 101 1001 1 001 1001
)
001 001 11 001 101 11
Random Bit Exchange Yet another variation on the chromosome-recombination
theme is the so-called random bit exchange. Here, the random-number generator
selects a user-specified number of locations, and then swaps the bits at these
locations as illustrated below.
11011001 10011101
)
00100111 01100011
Here, the second and the sixth bits (counting from the left) were swapped. Note
that nothing will happen if the leftmost bit is exchanged because it has the same
value in both chromosomes. The number of exchanged bits can vary but most
applications prefer the number to be much smaller than the chromosome’s length.
A common practice in realistic applications is to combine two or more recombi-
nation operators. For instance, the selected pair of parents will with 50% probability
be subjected to a 2-point chromosome, with 30% probability to a random bit
exchange, and with 20% probability there will be no recombination at all.
Inversion Whereas the recombination operators act on pairs of chromosomes,
other operators act on single specimens. One such operator is mutation; another
is inversion. In a typical implementation, the random-number generator returns two
integers that define two locations in the binary string (similarly as in the 2-point
crossover). Then, the substring between the two positions is inverted as shown
below.
110 110 01 ) 110 011 01
Note that the order of the zeros and ones in the substring between the third and
the seventh bit (counting from the left) was reversed. The location of the two points
determines how much inversion impacts the chromosome. If the two integers are
close to each other, say, 4 and 7, then only a small part of the chromosome is
affected.
In advanced implementations, inversion is used to supplement mutation. For
instance, the probability that a given bit is mutated can be set to 0.2% whereas
each chromosome may have a 0.7% chance to see its random substring inverted.
Similarly as with mutation, care has to be taken to make sure the inversion operator
is used rarely. Excessive use may destroy the positive contribution of recombination.
Inversion and Premature Degeneration Much more than mutation, inversion is
very good at extricating the genetic search from premature degeneration. To see
why, take a look at the following degenerated population.
16.6 Some Advanced Versions 321
0 1 0 0 0
0 1 0 0 1
0 1 0 0 0
0 1 0 0 0
Inverting the middle three bits of the first chromosome, and the last three bits of
the second chromosome will result in the following population:
0 0 0 1 0
0 1 1 0 0
0 1 0 0 0
0 1 0 0 0
The reader can see that the diversity has indeed increased. This observation
suggests a simple way to handle premature degeneration: just increase, for a while,
the frequency of inversions, and perhaps also that of mutations.
• Explain the differences between one-point crossover, two-point crossover, and
random bit exchange.
• What specific aspect makes the recombination operators different from the
mutation and inversion operators?
• How does inversion affect the genetic search?
16.6 Some Advanced Versions
The genetic algorithm is a versatile general framework with almost infinite possibil-
ities of variations. This section will introduce two interesting techniques.
A Note on the Lamarckian Alternative Computer programs are not constrained
by the limitations of biology. Very often, the engineer discards some of these
limitations, just as early aviators abandoned the idea of feathered wings. We have
already encountered one such violation when making some specimens “immortal,”
copying them into the new generation to make sure they would not be destroyed by
recombination and mutation. Let us now look at another deviation.
In the baseline genetic algorithm, new substrings come into being only as a result
of random processes during such operators as recombination or mutation. After this,
the genetic information remains unchanged throughout the specimen’s entire life.
One pre-Darwinian biologist, Jean-Baptiste Lamarck, suggested something more
flexible: in his view, evolution might be driven by the individuals’ needs. A giraffe
that keeps trying to reach the topmost leaves will stretch his neck that will thus
become longer. This longer neck is then passed on to the offspring. While the
lamarckian hypothesis is untenable in the realm of biology, it is not totally irrational
in other fields. For instance, by publishing a scientific paper, a researcher leaves to
posterity the knowledge acquired during his lifetime.
Lamarckian evolution is much faster than that of the classical darwinian process,
which is why we sometimes implement it in the genetic algorithm. The simplest
way to incorporate this concept in the general loop from Fig. 16.1 is to place the
“lamarckian” operator between the “wheel of fortune” and recombination. The task
for the operator is to improve the chromosome by adaptation. For instance, one
can ask what happens if a certain bit gets flipped-flopped by mutation. Whereas
mutation by itself is irreversible, we can add flexibility by explicitly testing what
happens when the i bit is flipped-flopped, and then choose the better version.
Multi-Population Search One motivation for multi-population search has to do
with the many parameters the genetic algorithm depends on. Most of the time, the
engineer has to rely only on her experience. Alternatively, we may choose to subject
the same initial population to several parallel runs of the genetic algorithm, each
with its own mutation frequency, with or without inversion, with a different mixture
of recombination operators, or with a modified fitness function. Among the many
alternatives, some will reach the solution faster than the others.
The reader will recall having encountered multi-population search in the section
that discussed the threat of premature degeneration. In that particular context, the
suggestion was to let two or more populations evolve in relative isolation that is
disrupted by occasional interbreeding. Note that this interbreeding may not be easy
to implement if each population uses a different way of chromosome definition as
suggested in the previous paragraphs. In that case, the programmer has to implement
a special program module for the conversion from one encoding to another.
Strings of Numbers, Strings of Symbols Chromosomes do not have to be binary
strings; they can consist of numbers, or characters. The same recombination
operators as before can then be used, though mutation may call for creativity.
Perhaps the most common kind of mutation in numeric strings is to use “noise”
superimposed on some (or all) of the chromosome’s “genes.” For instance, if all
locations contain numbers from the interval Œ0; 100, then the noise can be modeled
as a random number from Œa; a where a is a user-set parameter that plays here a
role similar to that of mutation frequency in binary strings. Here is how it can work:
Before mutation 10 22 17 42 16
The “noise” 3 1 2
After mutation 10 19 18 40 16
16.6 Some Advanced Versions 323
Fig. 16.5 A tree

representation of a candidate
expression from the “pies”
domain
shape = circle crust−shade = grey

crust−size = thick
The situation is slightly different if the chromosomes have the form of strings
of symbols. Here, mutation can replace a randomly selected symbol in the chromo-
some with another symbol chosen by the random-number generator. For instance,
when applied to chromosome [d s r d w k l], the mutation can change from
r to s the third symbol from the left, the resulting chromosome being [d s s d
w k l].
Also possible are “mixed” chromosomes where some locations are binary, others
numeric, and yet others symbolic. Here, mutation is usually implemented as a
combination of the individual approaches. For instance, the program selects a
random location in the chromosome, determines whether the location is binary,
numeric, or symbolic, and then applies the appropriate type of mutation.
Chromosomes Implemented as Tree Structures In some applications, strings of
bits, numbers, or symbols are inadequate; a tree-structure may then be more flexible.
This, for instance, is the case of classifiers in the form of logical expressions—see
the example in Fig. 16.5 where the following expression is represented by a tree-like
chromosome.
(shape=circle ^ crust-size=thick) _ : crust-shade=gray
The expression consists of attributes, the values of these attributes, and the logical
operators of conjunction, disjunction, and negation. Note how naturally this is cast in
the tree structure. The internal nodes represent the logical operations and the leaves
contain the attribute-value pairs Recombination swaps random subtrees. Mutation
can affect the leaves: either attribute names or attribute values or both. Another
possibility for mutation is occasionally to replace ^ with _ or the other way round.
Special attention has to be paid to the way the initial population is generated. The
programmer has to make sure that the population already contains some promising
expressions. One possibility is to create a group of random expressions and to
insert in it the descriptions of the positive examples. The survival function (to be
maximized) can be defined as the classification accuracy on the training set.
• What is the difference between the darwinian and the lamarckian evolution
processes? Which of them is faster?
• What weakness is remedied by the multi-population genetic algorithm? In what
way do multiple populations address this problem?
• How would you implement the mutation operator if the chromosome is a “mixed”
string of bits, numeric values, and symbols?
• How would you implement the recombination and mutation operators in domains
where chromosomes have the form of tree data structures?
16.7 Selections in k-NN Classifiers
Let us now illustrate a possible application of the genetic algorithm on a realistic

problem from the field of machine learning.
Attribute Selection and Example Selection The reader knows that the success of
the k-NN classifier depends on the quality of the stored examples and also on the
choice of the attributes to describe these examples. The problem of choosing the
right examples and attributes is easily cast in the search paradigm. For instance, the
initial state can be defined as the complete set of examples, and the complete set
of attributes; the search operators will remove examples and/or attributes; and the
evaluation function (whose value is to be minimized) will be defined as the error rate
reached by the 1-NN rule as measured on an independent set of testing examples.
Another possibility is to employ the genetic algorithm. In essence, we have to
decide how to represent the problem in terms of chromosomes, how to define the
fitness function, and the recombination and mutation operators. Then, we have to be
clear about how to interpret (and utilize) the result of the search.
Chromosomes to Encode the Problem A very simple approach will divide the
binary chromosome into two parts: each location in the first part corresponds to one
training example, and each location in the second part corresponds to one attribute.
If the value of a certain bit is 0, the corresponding example or attribute is ignored,
otherwise it is kept. The fitness function will be designed in a way that seeks to
minimize the number of 1s.
This solution may lead to impractically long chromosomes in domains where
the training set contains many examples: if the training set has ten thousand
examples, ten thousand bits would be needed. A better solution will then opt for
the more flexible variable-length scheme where each element in the chromosome
contains an integer that points to a training example or an attribute. The length of
16.7 Selections in k-NN Classifiers 325
chromosome 1 chromosome 2
SPECIMEN:
examples attributes
Fig. 16.6 Each specimen is described by two chromosomes, one representing examples and the
other representing attributes. Recombination is applied to each of them separately
the chromosome would be the number of relevant attributes plus the number of
representative examples. This mechanism is known as value encoding.
Interpreting the Chromosomes We must be sure to interpret the pairs of chro-
mosomes properly. For instance, the specimen [3,14,39],[2,4] represents a
training subset consisting of the third, the fourteenth, and the thirty-ninth training
example, described by the second and the fourth attribute. When such specimen is
used as a classifier, the system selects the examples determined by the first chromo-
some and describes them by the attributes determined by the second chromosome
(Fig. 16.6). The distances between vectors x D .x1 ; : : : xn / and y D .y1 ; : : : ; yn / are
calculated using the formula:
q
D.x; y/ D †niD1 d.xi ; yi / (16.1)
where d.xi ; yi / is the contribution of the ith dimension. For numeric attributes,
this contribution can be calculated by the usual formula for Euclidean distance,
d.xi ; yi / D .xi yi /2 ; for boolean attributes and for discrete attributes, we may
define d.xi ; yi / D 0 if xi D yi and d.xi ; yi / D 1 if xi ¤ yi .
The Fitness Function The next problem is how to quantify each individual’s
survival chances. Recall that we want to reduce the number of examples and
the number of attributes without compromising classification accuracy. These
requirements may contradict each other because, in noise-free domains, the entire
training set tends to give higher classification performance than a reduced set.
Likewise, removing attributes is hardly beneficial if each of them provides relevant
information.
The involved trade-offs therefore should be reflected in fitness-function param-
eters that give the user the chance to specify the concrete preferences. The
fitness function should make it possible to place emphasis either on maximizing
the classification accuracy or on minimizing the number of the retained training
examples and attributes. This requirement is expressed by the following formula
where ER is the number of training examples misclassified by the given specimen,
NE is the number of retained examples, and NA is the number of retained attributes:
f D 1=.c1 ER C c2 NE C c3 NA / (16.2)
Note that the fitness of a specimen is high if its error rate is low, if the set
of retained examples is small, and if many attributes have been eliminated. The
function is controlled by three user-set parameters, c1 ; c2 , and c3 , that weigh the
user’s preferences. For instance, if c1 is high, emphasis is placed on classification
accuracy. If c2 or c3 are high, emphasis is placed on minimizing the number of
retained examples and on minimizing the number of retained attributes, respectively.
Genetic Operators for This Application Parents are selected probabilistically.
In particular, the following formula is used to calculate the probability that the
specimen S’ will be chosen:
f .S0 /
Prob.S0 / D P (16.3)
f .S/
Here, f .S/ is the fitness of specimen S as calculated by Eq. (16.2). The denomina-
tor sums up the values of the fitness functions of all specimens in the population—
this makes the probabilities sum up to 1.
Once the pair of parents have been chosen, their chromosomes are recombined by
the two-point crossover. Since each specimen is defined by a pair of chromosomes,
each with a different meaning, we apply the recombination operator to each of them
separately. Let the length of one parent’s chromosome be denoted by N1 and let
the length of the other parent’s chromosome be denoted by N2 . Using the uniform
distribution, the algorithm selects one pair of integers from the closed interval Œ1; N1
and another pair of integers from the closed interval Œ1; N2 . Each of these pairs then
defines a substring in the respective chromosome (the first and the last locations
are included in the substring). The crossover operator then exchanges the substrings
from one of the parent chromosomes with the substrings of the other parent. Note
that, as each of these substrings can have a different size, the children’s lengths are
likely to be different from the parents’ lengths.
Graphical Illustration The principle is illustrated in Fig. 16.7 where the middle
parts of chromosomes A and B have been exchanged. Note how the lengths of A
and B are affected. The engineer has to decide whether to permit the situation where
the exchanged segments have size 0; in the other extreme, a segment can represent
the entire parent.
The mutation operator should prevent premature degeneration of the population
and make sure the population represents a representative part of the search space.
before crossover after crossover

A A’
B B’
Fig. 16.7 The two-point crossover operator creates the children by exchanging randomly selected
substrings in the parent chromosomes
One possibility is to select, randomly, a pre-specified percentage of the locations

in the newly created population and to add to each of them a random integer
generated separately for the location. The result is then taken modulo the number of
examples/attributes. Let the original number of examples/attributes be 100 and let
the location selected for mutation contains be 95. If the randomly generated integer
is 22, then the value after mutation is .95 C 22/ mod 100 D 17.
• What can be accomplished by choosing the best attributes and the most represen-
tative examples?
• What are the advantages of using two chromosomes instead of just one?
• How does the chosen fitness function reflect the competing requirements of small
sets of attributes and examples versus high classification accuracy?
• Why did we use a recombination operator that exchanges substrings of different
lengths? How was mutation carried out?
• The genetic algorithm, inspired by the Darwinian evolution, is a popular

alternative to classical artificial-intelligence search techniques. The simplest
implementation works with binary strings.
• The algorithm subjects a population of individuals to three essential operations:
fitness-function based survival, recombinations of pairs of chromosomes, and
mutation. Also inversion of a substring is sometimes used.
• One of the frequently encountered problems in practical applications of the
genetic algorithm is a population’s premature degeneration. One way of detecting
it is to consider the diversity of the chromosomes in the population. One solution
will add artificially created chromosomes to the population. Also the inversion
operator is useful, here.
• Alternative implementations of the genetic algorithm use strings of numbers,
symbols, mixed strings, or even tree structures.
• The chapter illustrated the practical use of the genetic algorithm using a simple
problem from the field of nearest-neighbor classifiers.
Historical Remarks The idea to cast the principle of biological evolution in the
form of the genetic algorithm is due to Holland [37], although some other authors
suggested something similar a little earlier. Among these, perhaps Rechenberg
[80] deserves to be mentioned, while Fogel et al. [29] should be credited with
pioneering the idea of genetic programming. The concrete way of applying the
genetic algorithm to selections in the k-classifier is from Rozsypal and Kubat [82].
Exercises
1. Hand-simulate the genetic algorithm with a pencil and paper in a similar way
as in Table 16.2. Use a fitness function of your own choice, a different initial
population, and the random points for a one-point crossover. Then repeat the
exercise with the two-point crossover.
1. Explain how different population sizes may affect the number of generations
needed to reach a good solution. Elaborate on the relation of population size to
the problem of premature degeneration. Discuss also the effect of the shape of
the fitness function.
2. What types of search problems are likely to be more efficiently addressed by the
genetic algorithm than by classical search algorithms?
3. Identify concrete engineering problems (other than those in the previous text)
appropriate for the genetic algorithm. Suggest problems where the chromosomes
are best represented by binary or numeric strings, and suggest problems where
trees are more appropriate.
4. Name some differences between natural evolution and its computer model.
Speculate on whether more inspiration can be taken from nature. Where do you
think are the advantages of the computer programs as compared to biological
evolution?
1. Implement the baseline genetic algorithm to operate on binary-string chromo-

somes. Make sure you have separate modules for the survival function, the wheel
of fortune, recombination, and mutation, and that these modules are sufficiently
general to enable easy modifications.
2. Create the initial populations for the “pies” and “circles” domains from Chap. 1
and use them as input to the program developed in the previous task. Note that, in
the case of the “circles” domain, you might have to consider a slight modification
of the original program so that it can handle numeric-string chromosomes.
3. For a domain of your choice, implement a few alternative mating strategies.
Run systematic experiments to find out which strategy will most quickly find
the solution. The speed can be measured by the number of chromosomes whose
fitness values have to be evaluated before the solution is found.
4. For a domain of your choice, experiment with alternative “cocktails” of different
recombination operators, and with different frequencies of recombinations,
mutations, and inversions. Plot graphs that show how the speed of search
(measured as in the previous task) depends on the concrete settings of these
parameters.
Chapter 17
Reinforcement Learning
The fundamental problem addressed by this book is how to induce a classifier

capable of determining the class of an object. We have seen quite a few techniques
that have been developed with this in mind. In reinforcement learning, though, the
task is different. Instead of induction from a set of pre-classified examples, the agent
“experiments” with a system, and the system responds to this experimentation with
rewards or punishments. The agent then optimizes its behavior, its goal being to
maximize the rewards and to minimize the punishments.
This alternative paradigm differs from the classifier-induction task to such an
extent that a critic might suggest that reinforcement learning should perhaps be
relegated to a different book, perhaps a sequel to this one. The wealth of available
material would certainly merit such decision. And yet, the author feels that this
textbook would be incomplete without at least a cursory introduction of the basic
ideas. Hence this last chapter.
17.1 How to Choose the Most Rewarding Action
To establish the terminology, and to convey some early understanding of what

reinforcement learning is all about, let us begin with a simplified version of the
task at hand.
N-Armed Bandit Figure 17.1 shows five slot machines. Each gives a different
average return, but we do not know how big these average returns are. If we want
to maximize our gains, we need to find out what these average returns are, and then
stick with the most promising machine. This is the essence of what machine learning
calls the problem of an N-armed bandit, alluding to the notorious tendency of the
slot machines to rob you of your money.

DOI 10.1007/978-3-319-63913-0_17
332 17 Reinforcement Learning
Fig. 17.1 The generic problem: which of the slot machines offers the highest average return?
In theory, this should be easy. Why not simply try each machine many times,
observe the returns, and then choose the one where these returns have been highest?
In reality, though, this is not a good idea. Too many coins may have to be wasted
before a reliable decision about the best machine can be made.
A Simple Strategy Mindful of the incurred costs, the practically minded engineer
will limit the experimentation, and make an initial choice based on just a few trials.
Knowing that this early decision is unreliable, she will not be dogmatic. She will
occasionally experiment with the other machines: what if some of them might
indeed be better? If yes, it will be quite reasonable to replace the “previously best”
with this new one. The strategy is quite natural. One does not have to be machine-
learning scientist to come up with something of this kind.
This then is the behavior that the reinforcement learning paradigm seeks to
emulate. In the specific case from Fig. 17.1, there are five actions to choose from.
The principle described above combines exploitation of the machine currently
believed to be the best, and the exploration of alternatives. Exploitation dominates;
exploration is rare. In the simplest implementation, the frequency of the exploration
steps is controlled by a user-specified parameter,
. For instance,
D 0:1 means that
the “best” machine (the one that appears best in view of previous trials) is chosen
90% of the time; in the remaining 10% cases, a chance is given to a randomly
selected other machine.
Keeping a Tally of the Rewards The “best action” is defined as the one that has
led to the highest average return.1 For each action, the learner keeps a tally of the
previous returns; and the average of these returns is regarded as this action’s quality.
For instance, let us refer to the machines in Fig. 17.1 by integers, 1; 2; 3; 4, and
5. Action ai then represents the choice of the i-th machine. Suppose the leftmost
machine was chosen three times, and these choices resulted in the following returns
r1 D 0; r2 D 9, and r3 D 3. The quality of this particular choice is then Q.a1 / D
.r1 C r2 C r3 /=3 D .0 C 9 C 3/=3 D 4.
To avoid the necessity to store the rewards of all previously taken actions, the
engineer implementing the procedure can take advantage of the following formula
where Qk .a/ is the quality of action a as calculated from k rewards, and rkC1 is the
.k C 1/st reward.
1
At this point, let us remark that the returns can be negative—”punishments,” rather.
17.1 How to Choose the Most Rewarding Action 333
Table 17.1 The algorithm for the

-greedy reinforcement learning strategy
Input: user-specified parameter
, e.g.,
D 0:1;
a set of actions, ai , and their initial value-estimates, Q0 .ai /;
for each action, ai , let ki D 0 (the number of times the action has been taken);
1. Generate a random number, p 2 .0; 1/, from the uniform distribution.

2. If p
, choose the action with the highest value (exploitation).
Otherwise, choose a randomly selected other action (exploration).
3. Denote the action chosen in the previous step by ai .
Observe the reward, ri .
4. Update the value of ai using the following formula:
1
Q.ai / D Q.ai / C Œri Q.ai /
ki C 1
5. Set ki D ki C 1 and return to 1.
1
QkC1 .a/ D Qk .a/ C ŒrkC1 Qk .a/ (17.1)
kC1
Thanks to this formula, it is enough to “remember” for each action only the
values of k and Qk .a/—these are all that is needed, together with the latest reward,
to update the action’s value at the .k C 1/st step.
The procedure just described is sometimes called the
-greedy strategy. For the
user’s convenience, Table 17.1 summarizes the algorithm in a pseudocode.
Initializing the Process To be able to use Formula (17.1), we need to start
somewhere: we need to set for each action its initial value, Q0 .ai /. An elegant
possibility is to choose a value well above any realistic single return. For instance,
if all returns are known to come from the interval Œ0; 10, the following will be
reasonable initial values: Q0 .ai / D 50.
At each moment, the system chooses, with (1
) probability, the action with the
highest value, breaking ties randomly. At the beginning, all actions have the same
chance of being taken. Suppose that ai is picked. In consequence of the received
reward, this action’s quality is then reduced using Formula (17.1). Therefore, when
the next action is to be selected, it will (if exploitation is to be used) have to be some
other action—whose value will then get reduced, too. Long story short, the reader
can see that initialization of all action values to the same big number makes sure
that, in the early stages of the game, all actions will be systematically experimented
with.
• Describe the
-greedy strategy to be used when searching for the best machine
in the N-armed bandit problem. Explain the meaning of actions and their values.
What is meant by exploitation and exploration?
• Describe the simple mechanism for maintaining the average rewards. How does
this mechanism update the action’s values?
• Why did this section recommend that the initial values, Q0 .ai /, of all actions
should be set to a multiple of the typical reward?
17.2 States and Actions in a Game
The example with slot machines is a simplification that has made it easy to explain
the basic terminology. Its main limitation is the existence of only one state in which
an appropriate action is to be selected.
In reality, the situation is more complicated than that. Usually, there are many
states, each with several actions to choose from. The essence can be illustrated on
the tic-tac-toe game.
The Tic-Tac-Toe Game The principle is shown in Fig. 17.2 for the elementary case
where the size of the playing board is three by three squares. Two players are taking
turns, one placing crosses on the board, the other one circles. The goal is to achieve
a line of three crosses or circles—either in a column or in a row or diagonally.
Who succeeds first, wins. If, in the situation on the left, it is the turn of the player
that plays with crosses, he wins by putting his cross in the bottom left corner. If,
conversely, it were his opponent’s turn, the opponent would prevent this by putting
there a circle.
States and Actions Each board-position represents a state. At each state, the player
is to choose a concrete action. Thus in the state depicted on the left, there are three
empty squares, and thus three actions to choose from (one of them winning). The
whole situation can be represented by a look-up table in which each state-action
o x x o x x
x o x o
o x o
Fig. 17.2 In tic-tac-toe, two players took turns at placing their crosses and circles. The winner is
the one who obtains a triplet in a line (vertical, horizontal, or diagonal)
17.2 States and Actions in a Game 335
pair has a certain value, Q.s; a/. Based on these values, the
-greedy policy decides
which action should be taken in the particular state. The action results in a reward,
r, and this reward is then used to update the value of the state-action pair by means
of Formula (17.1).
The most typical way of implementing the learning scenario is to let the program
play a long series of games with itself, starting with ad hoc choices for actions
(based on only the initial values of Q.s; a/), then gradually improving them until it
achieves very high playing strength.
The main problem is how to determine the rewards of the concrete actions. In
principle, three alternatives can be considered.
Episodic Formulation This is perhaps the simplest way of dealing with the
reward-assignment problem. A whole game is played. If it is won, then all state-
actions pairs encountered throughout the game by the learning agent are treated as
if they received reward 1. If the game is lost, they are treated as if they all received
reward 1.
The main weakness of this method is that it ignores the circumstance that not all
actions taken in a game have equally contributed to the final outcome. A player
may have lost only because of a single blunder that followed a long series of
excellent moves. In this case, it would of course be unfair, even impractical, to
punish the good moves. The same goes for the opposite: weak actions might actually
receive the reward only because the game happened to be eventually won thanks to
the opponent’s unexpected blunder. One can argue, however, that, in the long run,
these little “injustices” get averaged out because, most of the time, the winner’s
actions will be good.
The advantage of the episodic formulation is its simplicity.
Continuing Formulation The aforementioned problem with the episodic formu-
lation (the fact that it may punish a series of good moves on account of a single
blunder) might be removed under the assumption that we know how to determine
the reward right after each action. This is indeed sometimes possible; and even in
domain where this is not possible, one can often at least make an estimate.
Most of the time, however, an attempt to determine the reward for a given action
before the game ends is speculative—and thus misleading. This is why this approach
is rarely used.
Compromise: Discounted Returns This is essentially an episodic formulation
improved in a way that determines the rewards based on the length of the game. For
instance, the longer it has taken to win a tic-tac-toe game, the smaller the reward
should be. There is some logic, in this approach: stronger moves are likely to win
sooner. The way to implement this strategy is to discount the final reward by the
number of steps taken before the victory.
Here is how to formulate the idea more technically: Let rk denote the reward
obtained at the k-th trial and let 2 .0; 1/ is a user-set discounting constant. The
discounted return R is then calculated as follows:
Fig. 17.3 The task: keep the

pole upright by moving the
cart left or right
R D †1
kD1 rk
k
(17.2)
Note how the growing value of k decreases the coefficient by which rk is

multiplied. If the ultimate reward comes at the 10th step, and if “1” is the reward for
the winning game, then the discounted reward for D 0:9 is R D 0:910 1 D 0:35.
Illustration: Pole Balancing A good illustration of when the discounted return
may be a good idea is the pole-balancing problem shown in Fig. 17.3. Here, each
state is defined by such attributes as the cart location, the cart’s velocity, the pole’s
angle, and the velocity of the change in the pole’s angle. There are essentially two
actions to choose from: (1) apply force in the left-right direction or (2) apply force
in the right-left direction. However, a different amount of force may be used. The
simplest version of this task assumes that the actions can only be taken at regular
intervals, say, 0.2 s.
In this game, the longer the time that has elapsed before the pole falls, the greater
the perceived success, and this is why longer games should be rewarded more than
short games. A simple way to implement this circumstance is to reward each state
during the game with a 0, and the final fall with, say, r D 10. The discounted
return will then be R D 10 N where N is the number of steps before the pole has
fallen.
• Explain the difference between states and actions. What is the meaning of the
“value of the state-action pair”?
• When it comes to reward-assignment, what is the difference between the episodic
formulation and the continuing formulation?
• Discuss the motivation behind the idea of discounted returns. Give the precise
formula, and illustrate its use on the pole-balancing game.
17.3 The SARSA Approach 337
17.3 The SARSA Approach
The previous two sections introduced only a very simplified mechanism to deal with
the reinforcement-learning problem. Without going into details, let us describe here
a more popular approach that is known under the name of SARSA. The pseudocode
summarizing the algorithm is provided in Table 17.2.
Essentially, the episodic formulation with discounting is used. The episode
begins with selecting an initial state, s (in some domains, this initial state is
randomly generated). In a series of successive steps, actions are taken according
to the
-greedy policy. Each such action results in a new state, s0 , being reached,
and reward, r, being received. The same
-greedy policy is then used to choose the
next action, a0 (to be taken in state s0 ). After this, the quality, Q.s; a/, of the given
state-action pair is updated by the following formula:
Q.s; a/ D Q.s; a/ C ˛Œr C Q.s0 ; a0 / Q.s; a/ (17.3)
Here, ˛ is a user-set constant and is the discounting factor.

Note that the update of the state-action pair’s quality is based on the quintuple
.s; a; r; s0 ; a0 /. This is how the technique got its name.
• Describe the principle of the SARSA approach to reinforcement learning. Where
did this name come from?
Table 17.2 The SARSA algorithm—using the

-greedy strategy and the episodic formulation of
the task
Input: user-specified parameters
; ˛;
Initialized values of all action-value pairs, Q0 .si ; aj /;
for each state-action pair, si ; aj , initialize kij D 0;
1. Choose an initial state, s.

2. Choose action a using the
-greedy strategy from Table 17.1.
3. Take action a. This results in a new state, s0 , and reward, r.
4. In state s0 , choose action a0 using the
-greedy strategy.
Update Q.s; a/ D Q.s; a/ C ˛Œr C Q.s0 ; a0 / Q.s; a/
5. Let s D s0 and a D a0 .
If s is a terminal state, start a new episode by going to 1; otherwise, go to 3.
• Unlike the classifier-induction problems from the previous chapters, reinforce-

ment learning assumes that an agent learns from direct experimentation with a
system it is trying to control.
• In the greatly simplified formalism of the N-armed bandit, the agent seeks to
identify the most promising action—the one that offers the highest average
returns. The simplest practical implementation relies on the so-called
-greedy
policy.
• More realistic implementations of the task assume the existence of a set of states.
For each state, the agent is to choose from a set of alternative actions. The choice
can be made by the
-greedy policy that relies on the qualities of the state-action
pairs, Q.s; a/.
• The problem of assigning the rewards to the state-action pairs can be addressed
by its episodic formulation, by continuing formulation, or by episodic formula-
tion with discounting.
• Of the more advanced approaches to reinforcement learning, the chapter briefly
mentioned the SARSA method.
Historical Remarks One of the first systematic treatments of the “bandit” problem
was offered by Bellman [3] who, in turn, was building on some earlier work still.
Importantly, the same author later developed the principle of dynamic programming
that can be regarded as a direct precursor to reinforcement learning [4]. The basic
principles of reinforcement learning probably owe most for their development to
Sutton [87].
Exercises
1. Calculate the number of state-action pairs in the tic-tac-toe example from

Fig. 17.2.
1. This chapter is all built around the idea of using the

-greedy policy. What do
you think are the limitations of this policy? Can you suggest how to overcome
them?
2. The principles of reinforcement learning have been explained using some very
simple toy domains. Can you think of an interesting real-world application? The
main difficulty will be how to cast the concrete problem into the reinforcement-
learning formalism.
3. How many episodes might be needed to solve the simple version of the tic-tac-toe
game shown in Fig. 17.2?
1. Write a computer program that implements the N-armed bandit as described in

Sect. 17.1.
2. Consider the maze-problem illustrated in Fig. 17.4. The task is to find the shortest
path from the starting point, S, to the goal, G. A computer can use the principles
of reinforcement learning to learn this shortest path based on great many training
runs.
Suggest the data structures to capture the states and actions of this game. Write
a computer program that relies on the episodic formulation and the
-greedy
policy when addressing this task.
Fig. 17.4 The agent starts at

S; the task is to find the
shortest path to G
G
Bibliography
1. Ash, T. (1989). Dynamic node creation in backpropagation neural networks. Connection

Science: Journal of Neural Computing, Artificial Intelligence, and Cognitive Research, 1,
365–375.
2. Ball, G. H. & Hall, D. J. (1965). ISODATA, a novel method of data analysis and clasification.
Technical Report of the Standford University, Stanford, CA
3. Bellman, R. E. (1956). A problem in the sequential design of experiments. Sankhya, 16, 221–
229.
4. Bellman, R. E. (1957). Dynamic programming. Princeton: Princeton University Press.
5. Blake, C. L. & Merz, C. J. (1998). Repository of machine learning databases. Department
of Information and Computer Science, University of California at Irvine. www.ics.uci.edu/~
mlearn/MLRepository.html.
6. Blumer, W., Ehrenfeucht, A., Haussler, D., & Warmuth, M. K. (1989). Learnability and the
Vapnik-Chervonenkis dimension. Journals of the ACM, 36, 929–965.
7. Boutell, M. R., Luo, J., Shen, X., & Brown, C. M. (2004). Learning multi-label scene
classification. Pattern Recognition, 37, 1757–1771
8. Bower, G. H. & Hilgard, E. R. (1981). Theories of learning. Englewood Cliffs: Prentice-Hall.
9. Breiman, L. (1996). Bagging predictors. Machine Learning, 24, 123–140.
10. Breiman, L. (2001). Random forests. Machine Learning, 45, 5–32.
11. Breiman, L., Friedman, J., Olshen, R., & Stone, C. J. (1984). Classification and regression
trees. Belmont: Wadsworth International Group.
12. Broomhead, D. S. & Lowe, D. (1988). Multivariable functional interpolation and adaptive
networks. Complex Systems, 2, 321–355.
13. Bryson, A. E. & Ho, Y.-C. (1969). Applied optimal control. New York: Blaisdell.
14. Chow, C. K. (1957). An optimum character recognition system using decision functions. IRE
Transactions on Computers, EC-6, 247–254.
15. Clare, A. & King, R. D. (2001). Knowledge discovery in multi-label phenotype data. In
Proceedings of the 5th European conference on principles of data mining and knowledge
discovery, PKDD’01, Freiburg, Germany (pp. 42–53)
16. Clark, P. & Niblett, R. (1989). The CN2 induction algorithm. Machine Learning, 3, 261–284.
17. Coppin, B. (2004). Artificial intelligence illuminated. Sudbury: Jones and Bartlett.
18. Cover, T. M. (1965). Geometrical and statistical properties of systems of linear inequalities
with applications in pattern recognition. IEEE Transactions on Electronic Computers, EC-14,
326–334.
19. Cover, T. M. (1968). Estimation by the nearest neighbor rule. IEEE Transactions on
Information Theory, IT-14, 50–55.

DOI 10.1007/978-3-319-63913-0
342 Bibliography
20. Cover, T. M. & Hart, P. E. (1967). Nearest neighbor pattern classification. IEEE Transactions
on Information Theory, IT-13, 21–27.
21. Dasarathy, B. V. (1991). Nearest-neighbor classification techniques. Los Alomitos: IEEE
Computer Society Press.
22. Dietterich, T. (1998). Approximate statistical tests for comparing supervised classification
learning algorithms. Neural Computation, 10, 1895–1923.
23. Dudani, S. A. (1975). The distance-weighted k-nearest-neighbor rule. IEEE Transactions on
Systems, Man, and Cybernetics, SMC-6, 325–327.
24. Fayyad, U. M. & Irani, K. B. (1992). On the handling of continuous-valued attributes in
decision tree generation. Machine Learning, 8, 87–102.
25. Fisher, R. A. (1936). The use of multiple measurement in taxonomic problems. Annals of
Eugenics, 7, 111–132.
26. Fisher, D. (1987). Knowledge acquisition via incremental conceptual clustering. Machine
Learning, 2, 139–172.
27. Fix, E. & Hodges, J. L. (1951). Discriminatory analysis, non-parametric discrimination.
USAF School of Aviation Medicine, Randolph Field, TX, Project 21-49-004, Report 4,
Contract AF41(128)-3
28. Freund, Y. & Schapire, R. E. (1996). Experiments with a new boosting algorithm. In Machine
Learning: Proceedings of the Thirteenth International Conference, Bari (pp. 148–156).
29. Fogel, L. J., Owens, A. J., & Walsh, M. J. (1966). Artificial intelligence through simulated
evolution. New York: Wiley.
30. Friedman, J. H., Bentley, J. L., & Finkel, R. A. (1977). An algorithm for finding best matches
in logarithmic expected time. ACM Transactions on Mathematical Software, 3(3), 209–226.
31. Gennari, J. H., Langley, P., & Fisher, D. (1990). Models of incremental concept formation.
Artificial Intelligence, 40, 11–61.
32. Godbole, S. & Sarawagi, S. (2004). Discriminative methods for multi-label classification. In
H. Dai, R. Srikant, & C. Zhang (Eds.), Lecture Notes in Artificial Intelligence (Vol. 3056,
pp. 22–30). Berlin/Heidelberg: Springer.
33. Good, I. J. (1965). The estimation of probabilities: An essay on modern Bayesian methods.
Cambridge: MIT.
34. Gordon, D. F. & desJardin, M. (1995). Evaluation and selection of biases in machine learning.
Machine Learning, 20, 5–22.
35. Hart, P. E. (1968). The condensed nearest neighbor rule. IEEE Transactions on Information
Theory, IT-14, 515–516.
36. Hellman, M. E. (1970). The nearest neighbor classification rule with the reject option. IEEE
Transactions on Systems Science and Cybernetics, 6, 179–185.
37. Holland, J. H. (1975). Adaptation in natural and artificial systems. Ann Arbor: University of
Michigan Press.
38. Holte, R. C. (1993). Very simple classification rules perform well on most commonly used
databases. Machine Learning, 11, 63–90.
39. Hunt, E. B., Marin, J., & Stone, P. J. (1966). Experiments in induction. New York: Academic
Press.
40. Katz, A. J., Gately, M. T., & Collins, D. R. (1990). Robust classifiers without robust features.
Neural Computation, 2, 472–479.
41. Kearns, M. J. & Vazirani, U. V. (1994). An introduction to computational learning theory.
Cambridge, MA: MIT Press.
42. Kodratoff, Y. (1988). Introduction to machine learning. London: Pitman.
43. Kodratoff, Y. & Michalski, R. S. (1990). Machine learning: An artificial intelligence approach
(Vol. 3). San Mateo: Morgan Kaufmann.
44. Kohavi, R. (1997). Wrappers for feature selection. Artificial Intelligence, 97(1–2), 273–324.
45. Kohonen, T. (1982). Self-organized formation of topologically correct feature maps. Biologi-
cal Cybernetics, 43, 59–69
46. Kohonen, T. (1990). The self-organizing map. Proceedings of the IEEE, 78(9), 1464–1480.
Bibliography 343
47. Koller, D. & Sahami, M. (1997). Hierarchically classifying documents using very few words.
In Proceedings of the 14th International conference on machine learning, ICML’07, San
Francisco, USA (pp. 170–178)
48. Kononenko, I., Bratko, I., & Kukar, M. (1998). Application of machine learning to medical
diagnosis. In R. Michalski, I. Bratko, & M. Kubat (Eds.), Machine learning and data mining:
Methods and applications. Chichester: Wiley.
49. Kubat, M. (1989). Floating approximation in time-varying knowledge bases. Pattern Recog-
nition Letters, 10, 223–227.
50. Kubat, M., Pfurtscheller, G., & Flotzinger D. (1994). AI-based approach to automatic sleep
classification. Biological Cybernetics, 79, 443–448.
51. Kubat, M., Holte, R., & Matwin, S. (1997). Learning when negatives examples abound. In
Proceedings of the European conference on machine learning (ECML’97), Apr 1997, Prague
(pp. 146–153).
52. Kubat, M., Holte, R., & Matwin, S. (1998). Detection of oil-spills in radar images of sea
surface. Machine Learning, 30, 195–215.
53. Kubat, M., Koprinska, I., & Pfurtscheller, G. (1998). Learning to classify medical signals. In
R. Michalski, I. Bratko, & M. Kubat (Eds.), Machine learning and data mining: Methods and
applications. Chichester: Wiley.
54. Littlestone, N. (1987). Learning quickly when irrelevant attributes abound: A new linear
threshold algorithm. Machine Learning, 2, 285–318.
55. Lewis, D. D. & Gale, W. A. (1994). A sequential algorithm for training text classifiers.
In Proceedings of the 17th annual international ACM SIGIR conference on research and
development in information retrieval (SIGIR’94), Dublin (pp. 3–12).
56. Louizou, G. & Maybank, S. J. (1987). The nearest neighbor and the bayes error rates. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 9, 254–262.
57. McCallum, A. (1999). Multi-label text classification with a mixture model trained by EM. In
Proceedings of the workshop on text learning (AAAI’99) (pp. 1–7).
58. McQueen, J. (1967). Some methods for classification and analysis of multivariate obser-
vations. In Proceedings of the 5th Berkeley symposium on mathematical statistics and
probability, Berkeley (pp. 281–297).
59. Michalski, R. S. (1969). On the quasi-minimal solution of the general covering problem. In
Proceedings of the 5th international symposium on information processing (FCIP’69), Bled,
Yugoslavia (Vol. A3, pp. 125–128).
60. Michalski, R. S. & Tecuci, G. (1994). Machine learning: A multistrategy approach. Palo Alto:
Morgan Kaufmann.
61. Michalski, R. S., Carbonell, J. G., & Mitchell, T. M. (1983). Machine learning: An artificial
intelligence approach. Palo Alto: Tioga Publishing Company.
62. Michalski, R. S., Carbonell, J. G., & Mitchell, T. M. (1986). Machine learning: An artificial
intelligence approach (Vol. 2). Palo Alto: Tioga Publishing Company.
63. Michalski, R., Bratko, I., & Kubat, M. (1998). Machine learning and data mining: Methods
and applications. New York: Wiley.
64. Michell, M. (1998). An introduction to genetic algorithm. Cambridge, MA: MIT.
65. Mill, J. S. (1865). A system of logic. London: Longmans.
66. Minsky, M. & Papert, S. (1969). Perceptrons. Cambridge, MA: MIT.
67. Mitchell, T. M. (1982). Generalization as search. Artificial Intelligence, 18, 203–226.
68. Mitchell, T. M. (1997). Machine learning. New York: McGraw-Hill.
69. Mori, S, Suen, C. Y., & Yamamoto, K. (1992). Historical overview of OCR research and
development. Proceedings of IEEE, 80, 1029–1058.
70. Muggleton, S. & Buntine, W. (1988). Machine invention of first-order predicates by inverting
resolution. In Proceedings of the 5th international machine learning conference, Ann Arbor,
Michigan (pp. 339–352)
71. Murty, M. N. & Krishna, G. (1980). A computationally efficient technique for data clustering.
Pattern Recognition, 12, 153–158.
344 Bibliography
72. Neyman, J. & Pearson E. S. (1928). On the use and interpretation of certain test criteria for
purposes of statistical inference. Biometrica, 20A, 175–240.
73. Ogden, C. K. & Richards, I. A. (1923). The meaning of meaning (8th ed., 1946). New York:
Harcourt, Brace, and World.
74. Parzen E. (1962). On estimation of a probability density function and mode. Annals of
Mathematical Statistics, 33, 1065–1076.
75. Quinlan, J. R. (1979). Discovering rules by induction from large collections of examples. In
D. Michie (Ed.) Expert systems in the micro electronic age. Edinburgh: Edinburgh University
Press.
76. Quinlan, J. R. (1986). Induction of decision trees. Machine Learning, 1, 81–106.
77. Quinlan, R. (1990). Learning logical definitions from relations. Machine Learning, 5, 239–
266
78. Quinlan, J. R. (1993). C4.5: Programms for machine learning. San Mateo: Morgan Kauf-
mann.
79. Read, J., Pfahringer, B., Holmes, G., & Frank, E. (2011). Classifier chains for multi-label
classification. Machine Learning, 85, 333–359.
80. Rechenberg, I. (1973). Evolutionsstrategie: Optimierung technischer Systeme nach Princip-
ien der biologischen Evolution. Stuttgart: Frommann-Holzboog.
81. Rosenblatt, M. (1958). The perceptron: A probabilistic model for information storage and
organization in the brain. Psychological Review, 65, 386–408.
82. Rozsypal, A. & Kubat, M. (2001). Using the genetic algorithm to reduce the size of a nearest-
neighbor classifier and to select relevant attributes. In Proceedings of the 18th international
conference on machine learning, Williamstown (pp. 449–456).
83. Rumelhart, D. E. & McClelland, J. L. (1986). Parallel distributed processing. Cambridge:
MIT Bradford Press.
84. Russell, S. & Norvig, P. (2003). Artificial intelligence, a modern approach (2nd ed.).
Englewood Cliffs: Prentice Hall.
85. Schapire, R. E. (1990). The strength of weak learnability. Machine Learning, 5, 197–227.
86. Shawe-Taylor, J., Anthony, M., & Biggs, N. (1993). Bounding sample size with the Vapnik-
Chervonenkis dimension. Discrete Applied Mthematics, 42(1), 65–73.
87. Sutton, R. S. (1984). Temporal credit assignment in reinforcement learning. PhD Dissertation,
University of Massachusetts, Amherst.
88. Thrun, S. B. & Mitchell, T. M. (1995). Lifelong robot learning. Robotics and Automonous
Systems, 15, pp. 24–46.
89. Tomek, I. (1976). Two modifications of CNN. IEEE Transactions on Systems, Man and
Communications, SMC-6, 769–772.
90. Turney, P. D. (1993). Robust classification with context-sensitive features. In Proceedings
of the sixth international conference of industrial and engineering applications of artificial
intelligence and expert systems, Edinburgh (pp. 268–276).
91. Valiant, L. G. (1984). A theory of the learnable. Communications of the ACM, 27, 1134–1142.
92. Vapnik, V. N. (1992). Estimation of dependences based on empirical data. New York:
Springer.
93. Vapnik, V. N. (1995). The nature of statistical learning theory. New York: Springer.
94. Vapnik, V. N. & Chervonenkis, A. Y. (1971). On the uniform convergence of relative
frequencies of events to their probabilities. Theory of Probability and its Applications, 16,
264–280.
95. Werbos, P. (1974). Beyond regression: New tools for prediction and analysis in the behavioral
sciences. PhD thesis, Harvard University.
96. Whewel, W. (1858). History of scientific ideas. London: J.W. Parker.
97. Widmer, G. & Kubat, M. (1996). Learning in the presence of concept drift and hidden
contexts. Machine Learning, 23, 69–101.
Bibliography 345
98. Widrow, B. & Hoff, M. E. (1960). Adaptive switching circuits. In IRE WESCON convention
record, New York (pp. 96–104).
99. Wolpert, D. (1992). Stacked generalization. Neural Networks, 5, 241–259.
100. Wolpert, D. (1996). The lack of a priori distinctions between learning algorithms. Neural
Computation, 8, 1341–1390.
101. Zhang, M.-L. & Zhou, Z.-H. (2007). ML-KNN: A lazy learing approach to multi-label
learning. Pattern Recognition, 40, 2038–2048.
Index
A G
applications, 151, 167 gaussian function
attributes in Bayes, 33
continuous, 30, 38, 45–47, 84, 122, 145 in RBF networks, 107
discrete, 8, 22, 44, 137
irrelevant, 13, 49, 74, 75, 118, 144, 156
redundant, 13, 59, 118, 144, 156 I
selection, 204, 205, 324 imbalanced classes, 78, 194, 215, 225, 255
unknown, 202 interpretability, 114, 126
B
backpropagation, 98 L
bias, 67, 143, 191, 193 linear classifiers
in RBF networks, 108
perceptron, 69
C WINNOW, 73
clustering linearly-ordered classes, 207
hierarchical aggregation, 283
intercluster distance, 275, 284
k-means, 277, 281 M
normalization, 278 multi-label classification
principle, 273 binary relevance, 254
SOFM, 286 class aggregation, 263
context, 191, 199, 201 classifier chains, 256
nearest-neigbor classifiers, 253
neural networks, 252
D stacking, 258
decision trees
as classifiers, 113
converted to rules, 130
induction, 117 N
numeric, 122 nearest neighbor
pruning, 126 dangerous examples, 57
distance, 275 weighted, 55

DOI 10.1007/978-3-319-63913-0
348 Index
neural networks reinforcement

backpropagation, 97 sarsa, 337
MLP architecture, 100 states and actions, 334
MLP as classifiers, 91 rule induction
RBF networks, 106 predicates, 303
noise recursion, 303
in attributes, 14, 45, 57 rulesets, 130, 298
in class labels, 14 sequential covering, 300
normalization, 51, 278, 279, 289
P S
performance criteria search
Fˇ , 219 genetic, 309
error rate, 15 hill-climbing, 1, 5, 6, 8, 97, 309
macro-averaging, 265 similarity, 43
micro-averaging, 266 statistical evaluation, 241
precision, 215 statistical significance
recall, 215 margin or error, 239
sensitivity, 220 type I error, 243
specificity, 220 support vector machines
polynomial classifiers, 79 linear, 84
predicates RBF-based, 108
alternative search operators, 305
informal definition, 303
recursive rules, 303 T
probability, 19 time-varying classes, 200
pruning
decision tree, 127, 128, 184, 192
rules, 130, 131
V
voting
R plain, 174, 176
regression, 207 weighted majority, 179, 181

An Introduction To Machine Learning 2nd Ed

Uploaded by

Copyright:

Available Formats

An Introduction To Machine Learning 2nd Ed

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

An Introduction To Machine Learning 2nd Ed

Uploaded by

Copyright:

Available Formats

What are some of the main machine learning algorithms discussed in the book?

What are some of the main machine learning algorithms discussed in the book?

How does the book describe handling continuous attributes in Bayesian classifiers?

How does the book describe handling continuous attributes in Bayesian classifiers?

Miroslav

ISBN 978-3-319-63912-3 ISBN 978-3-319-63913-0 (eBook)

Library of Congress Control Number: 2017949183

© Springer International Publishing AG 2015, 2017

Printed on acid-free paper

This Springer imprint is published by Springer Nature

1 A Simple Machine-Learning Task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

4 Inter-Class Boundaries: Linear and Polynomial Classifiers . . . . . . . . . . . 65

8.6 Text Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

13.4 Another Possibility: Stacking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258

The bombshell exploded in 1983. Machine Learning: The AI Approach1 was

1.1 Training Sets and Classifiers

© Springer International Publishing AG 2017 1

Johnny does NOT like:

Table 1.1 The twelve training examples expressed in a matrix form

Attribute Vectors To be able to communicate the training examples to the

Problems with a Brute-Force Approach How does a machine find a classifier of

What Have You Learned?

1.2 Minor Digression: Hill-Climbing Search

Search Operators Search Strategy

Initial State Final State

Hill Climbing Final State:

Table 1.2 Hill-climbing search algorithm

What Have You Learned?

1.3 Hill Climbing in Machine Learning

e = 2/12 e = 6/12 e = 5/12

Fig. 1.4 Hill-climbing search in the “pies” domain

($3, 1.2lb) ($3, 1.2lb)

Hill Climbing in a Domain with Numeric Attributes

What Have You Learned?

1.4 The Induced Classifier’s Performance

Fig. 1.6 Pre-classified

What Have You Learned?

1.5 Some Difficulties with Available Data

because it can be calculated by subtracting date-of-birth from today’s date.

What Have You Learned?

1.6 Summary and Historical Remarks

2.5 b 1.4 c 2.3 d

representation; and William Whewel [96] advocated prototypical representations

1.7 Solidify Your Knowledge

Give It Some Thought

1. Write a program implementing hill climbing and apply it to the sliding-tiles

2.1 The Single-Attribute Case

Let us start with something so simple as to be unrealistic: a domain where each

© Springer International Publishing AG 2017 19

Fig. 2.1 The prior

of an example being positive given that filling-size=thick” is 37.5%—this

Applying Conditional Probability to Classification Importantly, the relative fre-

P.posjthick/ P.thick/ D P.thickjpos/ P.pos/

What Have You Learned?

2.2 Vectors of Discrete Attributes

Table 2.1 Illustrating the principle of Bayesian decision making

ex1 ex2 ex3 ex4 ex5 ex6 ex7 ex8

(note that P.posjthin/ C P.negjthin/ D P.posjthick/ C P.negjthick/ D 1)

Based on these results, we conclude that an example with filling-size=thin should

attribute is irrelevant, then the term .x2 y2 /2 is superfluous—and yet it affects,

Œ7; 4; 25; 5; 10