PATTERN

CLASSIFICATION
Second Edition

Richard O. Duda
Peter E . Hart
David G . Stork

A Wiley-Interscience Publication
JOHN WILEY & SONS, INC .
New York Chichester Weinheim Brisbane Singapore Toronto
CONTENTS

PREFACE xvii

INTRODUCTION 1

1.1 Machine Perception, 1

1 .2 An Example, 1
1 .2.1 Related Fields, 8
1.3 Pattern Recognition Systems, 9
1 .3.1 Sensing, 9
1 .3 .2 Segmentation and Grouping, 9
1 .3 .3 Feature Extraction, 11
1 .3 .4 Classification, 12
1 .3 .5 Post Processing, 13
1 .4 The Design Cycle, 14
1 .4.1 Data Collection, 14
1 .4.2 Feature Choice, 14
1 .4.3 Model Choice, 15
1 .4.4 Training, 15
1 .4.5 Evaluation, 15
1 .4.6 Computational Complexity, 16
1 .5 Learning and Adaptation, 16
1 .5.1 Supervised Learning, 16
1 .5.2 Unsupervised Learning, 17
1 .5.3 Reinforcement Learning, 17
1 .6 Conclusion, 17
Summary by Chapters, 17
Bibliographical and Historical Remarks, 18
Bibliography, 19

BAYESIAN DECISION THEORY 20

2.1 Introduction, 20
2.2 Bayesian Decision Theory-Continuous Features, 24
2.2.1 Two-Category Classification, 25
2.3 Minimum-Error-Rate Classification, 26
4°2.3.1 Minimax Criterion, 27

vn
Viii CONTENTS

*2.3.2 Neyman-Pearson Criterion, 28

2.4 Classifiers, Discriminant Functions, and Decision Surfaces, 29
2.4.1 The Multicategory Case, 29
2.4.2 The Two-Category Case, 30
2.5 The Normal Density, 31
2.5 .1 Univariate Density, 32
2.5 .2 Multivariate Density, 33
2.6 Discriminant Functions for the Normal Density, 36
2.6.1 Case 1 : o- 1, 36
%j = 2

2.6.2 Case 2: %i = %, 39
2.6.3 Case 3: = arbitrary, 41
%j

Decision Regions for Two-Dimensional

Gaussian Data, 41
*2.7 Error Probabilities and Integrals, 45
*2.8 Error Bounds for Normal Densities, 46
2.8.1 Chernoff Bound, 46
2.8.2 Bhattacharyya Bound, 47
I- ' Error Bounds for Gaussian Distributions, 48
2.8.3 Signal Detection Theory and Operating Characteristics, 48
2.9 Bayes Decision Theory-Discrete Features, 51
2.9.1 Independent Binary Features, 52
-, : Bayesian Decisions for Three-Dimensional
~ i i ~ 1 ;1 (= _'.

Binary Data, 53
*2.10 Missing and Noisy Features, 54
2.10.1 Missing Features, 54
2.10 .2 Noisy Features, 55
*2.11 Bayesian Belief Networks, 56
Belief Network for Fish, 59
*2.12 Compound Bayesian Decision Theory and Context, 62
Summary, 63
Bibliographical and Historical Remarks, 64
Problems, 65
Computer exercises, 80
Bibliography, 82

MAXIMUM-LIKELIHOOD AND BAYESIAN

PARAMETER ESTIMATION 84

3 .1 Introduction, 84
3.2 Maximum-Likelihood Estimation, 85
3.2.1 The General Principle, 85
3 .2.2 The Gaussian Case: Unknown 88
it,

3.2.3 The Gaussian Case: Unknown Ix and Y,, 88

3.2.4 Bias, 89
3.3 Bayesian Estimation, 90
3.3.1 The Class-Conditional Densities, 91
3.3.2 The Parameter Distribution, 91
3.4 Bayesian Parameter Estimation: Gaussian Case, 92
3.4.1 The Univariate Case : p(I.LI D), 92
3.4.2 The Univariate Case: p (x ID), 95
3.4.3 The Multivariate Case, 95
 CONTENTS N

3 .5 Bayesian Parameter Estimation: General Theory, 97

( Recursive Bayes Learning, 98
3.5.1 When Do Maximum-Likelihood and Bayes Methods Differ?, 100
3.5.2 Noninformative Priors and Invariance, 101
3.5.3 Gibbs Algorithm, 102
*3 .6 Sufficient Statistics, 102
3.6.1 Sufficient Statistics and the Exponential Family, 106
3 .7 Problems of Dimensionality, 107
3 .7.1 Accuracy, Dimension, and Training Sample Size, 107
3.7.2 Computational Complexity, 111 ,
3.7.3 Overfitting, 113
*3 .8 Component Analysis and Discriminants, 114
3.8.1 Principal Component Analysis (PCA), 115
3.8.2 Fisher Linear Discriminant, 117
3.8.3 Multiple Discriminant Analysis, 121
*3 .9 Expectation-Maximization (EM), 124
~;, oaniaYe 2 Expectation-Maximization for a 2D Normal Model, 126
3.10 Hidden Markov Models, 128
3.10.1 First-Order Markov Models, 128
3.10.2 First-Order Hidden Markov Models, 129
3.10.3 Hidden Markov Model Computation, 129
3.10.4 Evaluation, 131
, 5 :.o-q*~ Hidden Markov Model, 133
3.10.5 Decoding, 135
;: ~,r?rrrj~rs r~ HMM Decoding, 136
3.10.6 Learning, 137
Summary, 139
Bibliographical and Historical Remarks, 139
Problems, 140
Computer exercises, 155
Bibliography, 159

® NONPARAMETRIC TECHNIQUES 161

4.1 Introduction, 161

4.2 Density Estimation, 161
4.3 Parzen Windows, 164
4.3.1 Convergence of the Mean, 167
4.3.2 Convergence of the Variance, 167
4.3.3 Illustrations, 168
4.3.4 Classification Example, 168
4.3.5 Probabilistic Neural Networks (PNNs), 172
4.3.6 Choosing the Window Function, 174
4.4 Ic -Nearest-Neighbor Estimation, 174
4.4.1 k -Nearest-Neighbor and Parzen-Window Estimation, 176
4.4.2 Estimation of A Posteriori Probabilities, 177
4.5 The Nearest-Neighbor Rule, 177
4.5.1 Convergence of the Nearest Neighbor, 179
4.5.2 Error Rate for the Nearest-Neighbor Rule, 180
4.5.3 Error Bounds, 180
4.5.4 The k-Nearest-Neighbor Rule, 182
X CONTENTS

4.5.5 Computational Complexity of the k-Nearest-Neighbor Rule, 184

4.6 Metrics and Nearest-Neighbor Classification, 187
4.6.1 Properties of Metrics, 187
4 .6.2 Tangent Distance, 188
*4.7 Fuzzy Classification, 192
*4 .8 Reduced Coulomb Energy Networks, 195
4.9 Approximations by Series Expansions, 197
Summary, 199
Bibliographical and Historical Remarks, 200
Problems, 201
Computer exercises, 209
Bibliography, 213

5 LINEAR DISCRIMINANT FUNCTIONS 215

5.1 Introduction, 215
5.2 Linear Discriminant Functions and Decision Surfaces, 216
5.2.1 The Two-Category Case, 216
5.2.2 The Multicategory Case, 218
5.3 Generalized Linear Discriminant Functions, 219
5.4 The Two-Category Linearly Separable Case, 223
5.4.1 Geometry and Terminology, 224
5.4.2 Gradient Descent Procedures, 224
5.5 Minimizing the Perceptron Criterion Function, 227
5 .5.1 The Perceptron Criterion Function, 227
5 .5.2 Convergence Proof for Single-Sample Correction, 229
5.5.3 Some Direct Generalizations, 232
5.6 Relaxation Procedures, 235
5.6.1 The Descent Algorithm, 235
5.6.2 Convergence Proof, 237
5.7 Nonseparable Behavior, 238
5.8 Minimum Squared-Error Procedures, 239
5.8.1 Minimum Squared-Error and the Pseudoinverse, 240
i Constructing a Linear Classifier by Matrix
Pseudoinverse, 241
5 .8 .2 Relation to Fisher's Linear Discriminant, 242
5.8.3 Asymptotic Approximation to an Optimal Discriminant, 243
5 .8.4 The Widrow-Hoff or LMS Procedure, 245
5.8.5 Stochastic Approximation Methods, 246
5.9 The Ho-Kashyap Procedures, 249
5 .9.1 The Descent Procedure, 250
5 .9.2 Convergence Proof, 251
5 .9.3 Nonseparable Behavior, 253
5.9.4 Some Related Procedures, 253
*5.10 Linear Programming Algorithms, 256
5 .10.1 Linear Programming, 256
5 .10.2 The Linearly Separable Case, 257
5 .10.3 Minimizing the Perceptron Criterion Function, 258
*5,11 Support Vector Machines, 259
5 .11 .1 SVMTraining, 263
_ilopk ;2 SVM for the XOR Problem, 264
 CONTENTS A

5 .12 Multicategory Generalizations, 265

5.12.1 Kesler's Construction, 266
5.12.2 Convergence of the Fixed-Increment Rule, 266
5.12.3 Generalizations for MSE Procedures, 268
Summary, 269
Bibliographical and Historical Remarks, 270
Problems, 271
Computer exercises, 278
Bibliography, 281

MULTILAYER NEURAL NETWORKS 282

6.1 Introduction, 282
6.2 Feedforward Operation and Classification, 284
6.2.1 General Feedforward Operation, 286
6.2.2 Expressive Power of Multilayer Networks, 287
6 .3 Backpropagation Algorithm, 288
6.3.1 Network Learning, 289
6.3.2 Training Protocols, 293
6.3.3 Learning Curves, 295
6.4 Error Surfaces, 296
6.4.1 Some Small Networks, 296
6.4.2 The Exclusive-OR (XOR), 298
6.4.3 Larger Networks, 298
6.4.4 How Important Are Multiple Minima?, 299
6.5 Backpropagation as Feature Mapping, 299
6.5.1 Representations at the Hidden Layer-Weights, 302
6.6 Backpropagation, Bayes Theory and Probability, 303
6.6.1 Bayes Discriminants and Neural Networks, 303
6.6.2 Outputs as Probabilities, 304
*6.7 Related Statistical Techniques, 305
6.8 Practical Techniques for Improving Backpropagation, 306
6.8 .1 Activation Function, 307
6.8.2 Parameters for the Sigmoid, 308
6.8 .3 Scaling Input, 308
6.8 .4 Target Values, 309
6.8.5 Training with Noise, 310
6.8.6 Manufacturing Data, 310
6.8 .7 Number of Hidden Units, 310
6.8 .8 Initializing Weights, 311
6.8.9 Learning Rates, 312
6.8 .10 Momentum, 313
6.8.11 Weight Decay, 314
6.8 .12 Hints, 315
6.8.13 On-Line, Stochastic or Batch Training?, 316
6.8.14 Stopped Training, 316
6.8 .15 Number of Hidden Layers, 317
6.8.16 Criterion Function, 318
*6.9 Second-Order Methods, 318
6.9.1 Hessian Matrix, 318
6.9.2 Newton's Method, 319
Xii CONTENTS

6 .9.3 Quickprop, 320

6.9.4 Conjugate Gradient Descent, 321
( Conjugate Gradient Descent, 322
*6.10 Additional Networks and Training Methods, 324
6 .10.1 Radial Basis Function Networks (RBFs), 324
6.10 .2 Special Bases, 325
6.10.3 Matched Filters, 325
6.10.4 Convolutional Networks, 326
6.10.5 Recurrent Networks, 328
6.10.6 Cascade-Correlation, 329
6.11 Regularization, Complexity Adjustment and Pruning, 330
Summary, 333
Bibliographical and Historical Remarks, 333
Problems, 335
Computer exercises, 343
Bibliography, 347

STOCHASTIC METHODS 350

7.1 Introduction, 350
7.2 Stochastic Search, 351
7 .2.1 Simulated Annealing, 351
7.2.2 The Boltzmann Factor, 352
7.2.3 Deterministic Simulated Annealing, 357
7.3 Boltzmann Learning, 360
7.3.1 Stochastic Boltzmann Learning of Visible States, 360
7.3.2 Missing Features and Category Constraints, 365
7 .3.3 Deterministic Boltzmann Learning, 366
7.3.4 Initialization and Setting Parameters, 367
*7.4 Boltzmann Networks and Graphical Models, 370
7 .4.1 Other Graphical Models, 372
*7.5 Evolutionary Methods, 373
7.5.1 Genetic Algorithms, 373
7.5.2 Further Heuristics, 377
7.5.3 Why Do They Work?, 378
*7.6 Genetic Programming, 378
Summary, 381
Bibliographical and Historical Remarks, 381
Problems, 383
Computer exercises, 388
Bibliography, 391

NONMETRIC METHODS 394

8.1 Introduction, 394
8.2 Decision Trees, 395
8.3 CART, 396
8 .3.1 Number of Splits, 397
8 .3.2 Query Selection and Node Impurity, 398
8.3.3 When to Stop Splitting, 402
8.3.4 Pruning, 403
 CONTENTS Xiii

8 .3.5 Assignment of Leaf Node Labels, 404

VI : , ;r A Simple Tree, 404
~~, -ijrt ;, I

8.3.6 Computational Complexity, 406

8 .3.7 Feature Choice, 407
8.3.8 Multivariate Decision Trees, 408
8.3.9 Priors and Costs, 409
8.3.10 Missing Attributes, 409
aF;,cvmpdi_ '? Surrogate Splits and Missing Attributes, 410
8.4 Other Tree Methods, 411
8.4.1 ID3, 411
8.4.2 C4 .5, 411
8.4.3 Which Tree Classifier Is Best?, 412
*8.5 Recognition with Strings, 413
8.5.1 String Matching, 415
8.5 .2 Edit Distance, 418
8.5 .3 Computational Complexity, 420
8.5.4 String Matching with Errors, 420
8.5.5 String Matching with the "Don't-Care" Symbol, 421
8.6 Grammatical Methods, 421
8.6.1 Grammars, 422
8.6.2 Types of String Grammars, 424
1 , .11 -: i f ~I(: .s A Grammar for Pronouncing Numbers, 425
8.6.3 Recognition Using Grammars, 426
8.7 Grammatical Inference, 429
4 Grammatical Inference, 431
*8.8 Rule-Based Methods, 431
8.8 .1 Learning Rules, 433
Summary, 434
Bibliographical and Historical Remarks, 435
Problems, 437
Computer exercises, 446
Bibliography, 450

ALGORITHM-INDEPENDENT MACHINE LEARNING 453

9.1 Introduction, 453

9.2 Lack of Inherent Superiority of Any Classifier, 454
9.2.1 No Free Lunch Theorem, 454
1 No Free Lunch for Binary Data, 457
*9.2.2 Ugly Duckling Theorem, 458
9.2.3 Minimum Description Length (MDL), 461
9.2.4 Minimum Description Length Principle, 463
9.2.5 Overfitting Avoidance and Occam's Razor, 464
9.3 Bias and Variance, 465
9.3.1 Bias and Variance for Regression, 466
9.3 .2 Bias and Variance for Classification, 468
9.4 Resampling for Estimating Statistics, 471
9.4.1 Jackknife, 472
? Jackknife Estimate of Bias and Variance of the Mode, 473
9.4.2 Bootstrap, 474
9.5 Resampling for Classifier Design, 475
AV CONTENTS

9.5.1 Bagging, 475

9.5 .2 Boosting, 476
9.5 .3 Learning with Queries, 480
9.5.4 Arcing, Learning with Queries, Bias and Variance, 482
9.6 Estimating and Comparing Classifiers, 482
9.6.1 Parameterc Models, 483
9.6 .2 Cross-Validation, 483
9.6.3 Jackknife and Bootstrap Estimation of Classification Accuracy, 485
9.6 .4 Maximum-Likelihood Model Comparison, 486
9.6.5 Bayesian Model Comparison, 487
9.6 .6 The Problem-Average Error Rate, 489
9.6.7 Predicting Final Performance from Learning Curves, 492
9.6.8 The Capacity of a Separating Plane, 494
9.7 Combining Classifiers, 495
9.7.1 Component Classifiers with Discriminant Functions, 496
9.7.2 Component Classifiers without Discriminant Functions, 498
Summary, 499
Bibliographical and Historical Remarks, 500
Problems, 502
Computer exercises, 508
Bibliography, 513

UNSUPERVISED LEARNING AND CLUSTERING 517

10.1 Introduction, 517
10.2 Mixture Densities and Identifiability, 518
10.3 Maximum-Likelihood Estimates, 519
10.4 Application to Normal Mixtures, 521
10.4.1 Case l: Unknown Mean Vectors, 522
10.4.2 Case 2: All Parameters Unknown, 524
10.4.3 k-Means Clustering, 526
*10 .4.4 Fuzzy k-Means Clustering, 528
10.5 Unsupervised Bayesian Learning, 530
10.5.1 The Bayes Classifier, 530
10.5.2 Learning the Parameter Vector, 531
Exa 1 pie i Unsupervised Learning of Gaussian Data, 534
10.5.3 Decision-Directed Approximation, 536
10.6 Data Description and Clustering, 537
10.6.1 Similarity Measures, 538
10.7 Criterion Functions for Clustering, 542
10.7.1 The Sum-of-Squared-Error Criterion, 542
10.7.2 Related Minimum Variance Criteria, 543
10.7.3 Scatter Criteria, 544
?L. ;; ~~ipte ~? Clustering Criteria, 546
*10.8 Iterative Optimization, 548
10.9 Hierarchical Clustering, 550
10.9.1 Definitions, 551
10 .9.2 Agglomerative Hierarchical Clustering, 552
10.9.3 Stepwise-Optimal Hierarchical Clustering, 555
10 .9.4 Hierarchical Clustering and Induced Metrics, 556
10 .10 The Problem of Validity, 557
 CONTENTS XV

*10.11 On-line clustering, 559

10.11 .1 Unknown Number of Clusters, 561
10.11 .2 Adaptive Resonance, 563
10.11 .3 Learning with a Critic, 565
* 10.12 Graph-Theoretic Methods, 566
10.13 Component Analysis, 568
10.13 .1 Principal Component Analysis (PCA), 568
10.13 .2 Nonlinear Component Analysis (NLCA), 569
*10 .13 .3 Independent Component Analysis (ICA), 570
10.14 Low-Dimensional Representations and Multidimensional Scaling
(MDS), 573
10.14.1 Self-Organizing Feature Maps, 576
10.14.2 Clustering and Dimensionality Reduction, 580
Summary, 581
Bibliographical and Historical Remarks, 582
Problems, 583
Computer exercises, 593
Bibliography, 598

A~ MATHEMATICAL FOUNDATIONS 601

A.1 Notation, 601

A.2 Linear Algebra, 604
A.2.1 Notation and Preliminaries, 604
A.2.2 Inner Product, 605
A.2.3 Outer Product, 606
A.2.4 Derivatives of Matrices, 606
A.2.5 Determinant and Trace, 608
A.2.6 Matrix Inversion, 609
A.2.7 Eigenvectors and Eigenvalues, 609
A.3 Lagrange Optimization, 610
A.4 Probability Theory, 611
A.4.1 Discrete Random Variables, 611
A.4.2 Expected Values, 611
A.4.3 Pairs of Discrete Random Variables, 612
A.4.4 Statistical Independence, 613
A.4.5 Expected Values of Functions of Two Variables, 613
A.4.6 Conditional Probability, 614
A.4.7 The Law ofTotal Probability and Bayes Rule, 615
A.4 .8 Vector Random Variables, 616
AA.9 Expectations, Mean Vectors and Covariance Matrices, 617
A.4.10 Continuous Random Variables, 618
A.4.11 Distributions of Sums of Independent Random Variables, 620
A.4.12 Normal Distributions, 621
A.5 Gaussian Derivatives and Integrals, 623
A.5 .1 Multivariate Normal Densities, 624
A.5 .2 Bivariate Normal Densities, 626
A.6 Hypothesis Testing, 628
A.6.1 Chi-Squared Test, 629
A.7 Information Theory, 630
A.7.1 Entropy and Information, 630
XVi CONTENTS

A.7.2 Relative Entropy, 632

A.7.3 Mutual Information, 632
A.8 Computational Complexity, 633
Bibliography, 635