Online Classification of Nonstationary Data Streams

Mark Last1
Ben-Gurion University of the Negev
Department of Information Systems Engineering
Beer-Sheva 84105, Israel
Email: [email protected]

Abstract
Most classification methods are based on the assumption that the data conforms to
a stationary distribution. However, the real-world data is usually collected over certain
periods of time, ranging from seconds to years, and ignoring possible changes in the
underlying concept, also known as concept drift, may degrade the predictive performance
of a classification model. Moreover, the computation time, the amount of required
memory, and the model complexity may grow indefinitely with the continuous arrival of
new training instances. This paper describes and evaluates OLIN, an online classification
system, which dynamically adjusts the size of the training window and the number of
new examples between model re-constructions to the current rate of concept drift. By
using a fixed amount of computer resources, OLIN produces models, which have nearly
the same accuracy as the ones that would be produced by periodically re-constructing the
model from all accumulated instances. We evaluate the system performance on sample
segments from two real-world streams of non-stationary data.
Keywords. Classification, incremental learning, online learning, concept drift,
info-fuzzy networks.

Also with Department of Computer Science and Engineering, University of South Florida, USA

Intelligent Data Analysis

1. Introduction
The so-called Information Age has provided us with huge amounts of digital data,
which keep growing at unprecedented pace. Business information systems, Internet
servers, and telecommunication providers are producing and storing new data every day,
hour, and even minute. The reported rates include the maximum amount of 17,400 Web
page requests per minute at a single university campus [6], 500,000 transactions recorded
during less than one day by the Lycos search engine [1], and tens of millions of queries
sent every day to Yahoo! Search [25]. The accumulated data may contain some valuable
information for an organization that is storing it. Data mining is the core stage of the
KDD (Knowledge Discovery in Databases) process, which is aimed at identifying valid,
novel, potentially useful, and ultimately understandable patterns in data [9]. The
problems of knowledge discovery in online data include real-time monitoring of
manufacturing processes, prediction of stock prices, and intrusion detection in computer
networks.
Continuous streams of data pose new problems to classification methods of data
mining, like CART [4], ID3 [28], C4.5 [29], IFN [23], and many others. The common
approach of these methods is to store and process the entire set of training examples. The
growing amounts of training data increase the processing requirements of data mining
systems up to a point, where they either run out of memory, or their computation time
becomes prohibitively long. Furthermore, even if all the available examples can be
handled by the system, the patterns discovered by an algorithm in the data from the past,
may be hardly valid and useful for the new data obtained hours or even minutes later due

Mark Last

Intelligent Data Analysis

to unexpected changes in the data-generating process (e.g., a political event which affects
the stock prices).
Methods for extracting patterns from continuous streams of data are known as
incremental (online) learning algorithms. The basic idea of incremental induction is that
upon receiving a new instance, it is much less expensive to update an existing model than
to build a new one. On the other hand, as indicated in [6], the incremental algorithms
suffer from several shortcomings, like high sensitivity to the order of training examples
and longer training times than the non-incremental (batch) methods. Pure incremental
methods consider every new instance, which may be impractical in environments, where
transactions arrive at the rate of thousands per second. The algorithms for purely
incremental learning include WINNOW [21], COBWEB [10], ITI [31], and the stochastic
gradient descent algorithm for training a neural network (see [27]). Most of these
methods are focused on efficient ways of inducing a classification model from stationary
data streams.
The general problem of learning a drifting concept that is a concept which
changes over time is studied by Helmbold and Long in [12]. The rate of drift is defined
there as the probability that the target function disagrees over two successive examples.
Disagreements are minimized by an algorithm, which is polynomial in the sample size,
given that the rate of drift is bounded. Thus, for high-volume non-stationary data
streams, where the actual rate of drift is unknown in advance, the run time of the
algorithm may grow indefinitely.
Widmer and Kubat [32] describe a family of purely incremental algorithms for
learning in the presence of drift. These algorithms use a simple representational
Mark Last

Intelligent Data Analysis

framework, called FLORA, which stores descriptions of positive, negative, and noisy
examples in separate description sets. One of the incremental algorithms, FLORA2,
maintains a dynamically adjustable window of the latest training examples. Whenever a
concept drift is suspected, due to a drop in predictive accuracy or an explosion in the
number of descriptions, the window size is decreased, by discarding the oldest examples.
If the concept appears to be stable, the window size is left unchanged. As long as the
presence of drift is uncertain, no examples are forgotten, thus incrementally increasing
the window size. According to [32], this window adjustment strategy may efficiently
detect radical changes in the underlying concept, subject to a relatively low rate of
change. The FLORA algorithms also assume a limited rate of data arrival, since they
process one example at a time.
A recent paper by Domingos and Hulten [6] deals directly with the problem of
mining high-speed streams of data. Their data mining system, called VFDT, builds
decision trees from symbolic attributes by using sub-sampling of the entire data stream
generated by a stationary process. A similar assumption of stationary concepts is used by
the incremental method of Fan et al. [7]. The sample size is determined in VFDT from
distribution-free Hoeffding bounds. A new version of the VFDT system, called CVFDT
[16], learns decision trees from continuously changing data streams by repeatedly
applying the VFDT algorithm to a sliding window of fixed size. CVFDT is aimed at
detecting only one type of concept drift at the node level of the tree: namely, the
importance of the current input attribute vs. other attributes. The algorithm grows an
alternative subtree for each attribute having a relatively high information gain and
replaces the old subtree when a new one becomes more accurate.

Mark Last

Intelligent Data Analysis

Harries and Sammut [11] have developed an off-line meta-learning system (based
on C4.5) for partitioning the data stream into a set of time-dependent "conceptual
clusters." The off-line approach is aimed at analyzing concept drifts in historic data
rather than at the online detection of evolving changes. A strategy for dynamically
updating a linear regression classifier by using a linear model of dynamic behavior is
presented in [17]. The linear approach of[17] is computationally efficient, but it
obviously restricts the search space by eliminating nonlinear concepts and nonlinear
(especially, abrupt) patterns of concept drift.
Time-dependent changes in the class distributions of rules induced from data can
be detected by the CD3 algorithm of Black and Hickey [3][13]. CD3 treats the timestamp as an additional input attribute in a decision tree. Consequently, paths where the
value of the time-stamp attribute refers to the old period(s) represent rules, which are out
of date. When the process is stable for a long period of time, the time-stamp attribute
should not appear in any path of the tree.
Closely associated with the problem of change detection is the task of discovering
the robust knowledge, which is unlikely to be affected by database changes. A Bayesian
Network model for evaluating robustness of database rules is described in [14]. The
network probabilities are estimated by an off-line procedure, which assumes stationarity
of the database transactions.
Another related area is change detection, including an increasingly important
problem of intrusion detection in computers and computer networks. Lane and Brodley
[18] suggest a compromise between purely batch and purely incremental learning for the
task of detecting abnormal (possibly hostile) behavior of a computer user: their algorithm
Mark Last

Intelligent Data Analysis

is trained on a batch of the most recent user transactions. However, the minimum batch
size (80 examples) makes the system vulnerable to short-term cyber attacks. According
to Fawcett and Provost [8], the most efficient method for detecting an intruder is the
profiling approach, where a model of normal (stationary) activity is built and then used to
alarm on significant deviations from the normal.
This paper proposes an online classification system, which uses an info-fuzzy
network [23], or IFN, as a base classifier. As shown in [19][23], the IFN method is able
to produce much more compact models than other decision-tree methods, like CART and
C4.5, while preserving nearly the same level of predictive accuracy. Moreover, it can also
be used as an efficient feature selection method [20]. The proposed system, called OLIN
for On-Line Information Network, adapts itself automatically to the rate of concept drift
in a non-stationary data stream by dynamically adjusting the size of the training window
and the rate of model update. The system does not impose any limitations on the rate, the
extent, or the type of change in the underlying concept. Like the batch version of the IFN
method, it can handle both discrete and continuous attributes. OLIN saves computer
resources by increasing the update cycle when the concept appears to be stable and it
shrinks the size of the training window, whenever a concept drift is detected. Thus,
OLIN can be applied to a time-changing data stream of arbitrary duration. The
cumulative accuracy of the models produced by OLIN tends to be higher than the
accuracy obtained with a fixed-size sliding window though it may be slightly lower than
the accuracy of an incremental system that does not forget any past examples.
This paper is organized as follows. In the next section, we provide a brief
overview of the IFN method and its features. The OLIN system is described in the
Mark Last

Intelligent Data Analysis

following section. We then present the empirical results of applying OLIN to real-world
streams of non-stationary data coming from two different domains: semiconductor
manufacturing and stock market. The paper concludes with a discussion of results and
future work.

2. Information Networks
2.1. Network Structure
IFN, or info-fuzzy network [23] is a tree-like classification model, which is
designed to minimize the total number of predicting attributes. Beyond classification,
this model can be used for the tasks of discretization [23], feature selection [20], rule
induction [19], and evaluation of data reliability [24]. An info-fuzzy network has the
following components:
1) I - a subset of input (predicting) attributes used by the model. Input attributes are
selected from the set C of candidate input attributes (available features).
2) |I| - total number of hidden layers (levels) in a network. Unlike the standard decision
tree structure used in CART [4], ID3 [28], and C4.5 [29], where the nodes of the same
tree level are independent of each other, all nodes of a given network layer are labeled
by the same input attribute associated with that layer. This is why the number of
network layers is equal to the number of input attributes. In layers associated with
continuous attributes, an information network uses multiple splits, which are identical
at all nodes of the corresponding layer. Most other decision-tree methods apply only
binary splits in each layer. The first layer in the network (Layer 0) includes only the
root node and is not associated with any input attribute.
Mark Last

Intelligent Data Analysis

3) Ll - a subset of nodes z in a hidden layer l. Each node represents an attribute-based

test, similarly to a standard decision tree. If a hidden layer l is associated with a
nominal input attribute, each outgoing edge of a non-terminal node corresponds to a
distinct value of an attribute. For continuous features, the outgoing edges represent
consecutive intervals obtained from the discretization process. If a node has no
outgoing edges, it is called a terminal node.
4) K - a subset of target nodes representing distinct values of the target (classification)
attribute. For continuous target attributes, the target nodes represent disjoint intervals
in the attribute range. A target layer is missing in the standard decision-tree
structure. The connections between terminal nodes and the nodes of the target layer
may be used for extracting information-theoretic rules from a network [19].
In Figure 1, a structure of a two-layered info-fuzzy network (based on two
selected input attributes) is shown. The first input attribute has three values, represented
by nodes no. 1,2, and 3 in the first layer, but only nodes no. 1 and 3 are split by the
network construction procedure described in sub-section 2.2 below. The second layer has
four nodes standing for the combinations of two values of the second input attribute with
two split nodes of the first layer. The target attribute has three values, represented by
three nodes in the target layer. New examples can be classified by an info-fuzzy network
in a similar way to standard decision trees: we start at the root node, test the attribute
associated with the first layer, then move along the network path corresponding to the
value of the first input attribute. The process continues until a terminal node is
encountered (nodes 1,1; 1,2; 2; 3,1; and 3,2 in the network of Figure 1), at which time an

Mark Last

Intelligent Data Analysis

example is labeled with a single predicted class having the maximum probability at the
node or with the probability distribution of classes.
1,1

1
2

0
Layer No. 0
(the root node)

1,2
3,1
3,2

Layer No. 1
Connection
Layer No. 2
(First input
Weights
(Second input
attribute)
attribute)

2
3
Target
Layer

Figure 1 Info-Fuzzy Network - Two-Layered Structure

The connectionist nature of the info-fuzzy network (each terminal node is
connected to every target node) resembles the topological structure of multi-layer neural
networks (see [27]), which also have input and output nodes and a variable number of
hidden layers. Consequently, we define our model as a network and not as a tree.
2.2. Network Construction Procedure
The network construction algorithm (called here IN for Information Network)
starts with defining the target layer and the root node representing an empty set of input
attributes. Unlike CART [4] and C4.5 [29], IFN is built only in one direction (top-down).
After the construction is stopped, there is no bottom-up post-pruning of the network
branches. As explained below, the network is pre-pruned by applying statistical
significance testing.
A node z can be split on an input attribute Ai only if the split provides a
statistically significant increase in the mutual information of z and the target attribute Ai.
Mutual information (see [5]) is an information-theoretic measure of association between
Mark Last

Intelligent Data Analysis

two random variables X and Y, which is defined as a decrease in the entropy of Y as a

result of knowing X (and vice versa). An increase in mutual information, also called
conditional mutual information [5] or information gain [28], of a candidate input attribute
Ai and the target attribute Ai, given a node z, is calculated by the following expression
(based on [5]):
M

MI ( A i ' ; A i / z ) =

1M

i '

j= 0

P (V

ij

;V

i'

j ' ; z ) log

j '= 0

P (V
P (V

i' j'

ij
i' j'

/ z)

/ z ) P (V

ij

/ z)

(1)

where
Mi / Mi - number of distinct values of the target attribute i /candidate input
attribute i.
P (Vij/ z) - an estimated conditional (a posteriori) probability of a value j of a
candidate input attribute i given a node z
P (Vij/ z) - an estimated conditional (a posteriori) probability of a value j of the
target attribute i given a node z.
P (Vijij/ z) - an estimated conditional (a posteriori) probability of a value j of a
candidate input attribute i and a value j of the target attribute i given a node z.
P (Vij; Vij; z) - an estimated joint probability of a value j of the target attribute i, a
value j of a candidate input attribute i and a node z.
Conditional mutual information measures the benefit of adding an input attribute
to the information network. If the input and the target attributes are conditionally
independent given a node z, their conditional joint probability should be equal to the
product of their individual conditional probabilities. This makes the logarithmic terms in
Equation (1) equal to zero. On the other hand, if the knowledge of an input value either
Mark Last

10

Intelligent Data Analysis

increases, or decreases the conditional probability of a target value, the corresponding

summation term in (1) becomes either positive, or negative respectively.
If a tested attribute is continuous, its values in Equation (1) correspond to
thresholds, which maximize an increase in mutual information. Prior to adding a new
layer, the algorithm re-computes the best threshold splits of each continuous attribute that
is not in the network. More details on discretizing and selecting continuous attributes are
provided in [20] and [23].
The statistical significance of the estimated conditional mutual information, is
evaluated by using the likelihood-ratio statistic (based on [1]):
G2 (Ai ; Ai / z) = 2(ln2) E*(z) MI (Ai ; Ai / z)

(2)

Where E*(z) is the number of records associated with the node z

The Likelihood-Ratio Test [30] is a general-purpose method for testing the null
hypothesis H0 that two discrete random variables are statistically independent. As can be
seen from Equation (1), independence of two attributes implies that their expected mutual
information is zero. If H0 holds, then the likelihood-ratio test statistic G2 (Ai; Ai / z) is
distributed as chi-square with (NIi (z) - 1)( NTi (z) - 1) degrees of freedom, where
NIi(z) is the number of values of a candidate input attribute i at node z and NT i (z) is the
number of values of the target attribute i at node z (based on [30]). Thus, MI (Ai ; Ai / z)
is considered statistically significant if H0 can be rejected at the significance level :
G2 (Ai ; Ai / z) 2 ((NIi (z) - 1)( NTi (z) - 1))

(3)

The default value of used by the IN algorithm is 0.1%. We have found

empirically that in most datasets, higher values of tend to decrease the generalization
performance of the model.
Mark Last

11

Intelligent Data Analysis

At each iteration, the algorithm builds a new hidden layer by choosing an input
attribute (either discrete, or continuous), which provides the maximum significant
increase in mutual information relative to the previous layer. The nodes of a new layer
are defined for a Cartesian product of split nodes of the previous layer and the values of a
new input attribute. The chain rule of the information theory (see [5]) implies that the
mutual information between an info-fuzzy network and the target attribute is equal to the
sum of drops in conditional entropy (information gains) across all hidden layers.

If

there is no candidate input attribute significantly increasing the mutual information, the
network construction is stopped and the algorithm outputs the final network structure.

3. The OLIN System

3.1. Algorithm Description
The OLIN (On-Line Information Network) system is a wrapper for the
Information Network (IN) algorithm presented in the previous section. OLIN is receiving
a continuous stream of data examples. The system repeatedly applies the IN algorithm to
a sliding window of training examples and it dynamically adapts the size of the training
window and the frequency of model re-construction to the current rate of concept drift.
The OLIN system is limited neither by the overall duration of a data stream, nor by the
cumulative number of examples arrived. At each point in time, the purpose of the system
is to predict a correct class for the next arriving example by using a current classification
model. We assume that immediately afterwards the correct classification becomes
available to the wrapper. This assumption, used by many incremental learning methods

Mark Last

12

Intelligent Data Analysis

[32], is reasonable as long as the delay between receiving an example and receiving its
class is negligible with respect to the time between the arrivals of successive examples.
The general architecture of the OLIN system is presented in Figure 2. The system
has three modules: the Learning Module, which implements the IN algorithm to produce
an info-fuzzy network; the Classification Module, which uses the current network to
classify the incoming examples; and the Meta-Learning Module, which controls the
operation of the Learning Module. In the sample data stream of Figure 2, each one of V0
examples in the validation interval [t2, t3] is classified by a model induced from T0
examples of the training interval [t0, t2]. The number of examples in the training and the
validation intervals do not have to be equal. At the time t3 the network is re-constructed
by the Learning Module using T1 examples from the training interval [t1, t3] and
subsequently applied to V1 examples in the validation interval [t3, t4]. We assume here
that the first example in the interval [t3, t4] arrives after the network construction has been
completed. In massive data streams, examples that arrive in the process of network
construction may be classified by a partially constructed model using the anytime nature
of the IN algorithm (see [23]).
The Meta-Learning Module obtains as input the training and the validation
accuracy rates of the model, measured on the training and the validation intervals
respectively. It also gets the description of the model itself (selected attributes, entropy
information, etc.). Using the OLIN algorithm (see below), the module re-calculates the
size of the next training window (interval) and the number of validation examples to be
classified with the new model. The last validation example of the current window is
always the last training example of the next window. To classify every example in the
Mark Last

13

Intelligent Data Analysis

data stream exactly one time, the validation intervals ([t2, t3], [t3, t4], etc.) have to be
disjoint and consecutive. However, the training windows may overlap with each other
(but not with their respective validation intervals).
Meta-learning
Module
Window size

Learning
Module
Training
Data

Validation
accuracy

Training
accuracy

Update cycle

Classification
Model (IN)

Classification
Module
Validation
Data

t0

t1

t2

Current Window (T0 examples)

t3
V0
examples

t4
V1
examples

Data Stream

Next Window (T 1 examples)

Figure 2 OLIN System Architecture

The basic intuition behind the OLIN algorithm resembles the FLORA2
framework [32]: a narrow training window may be too small to identify any stable
concept with a high degree of confidence; on the other hand, a wide training window may
completely miss short-term changes in the underlying concept. Since the timing, the rate,
and the extent of concept drift in a data stream are not known in advance, the window
size should be adjusted dynamically in the process of data arrival. A good heuristic is to
shrink the training window to a minimum size, when a drift seems to occur, and increase
the window (up to a certain limit dictated by the data size and the available computer
resources) if stability is observed. OLIN uses the statistical significance of the difference

Mark Last

14

Intelligent Data Analysis

between the training and the validation accuracy of the current model as an indicator of
concept stability.
The purely incremental approach of FLORA2 does not seem appropriate for
massive data streams, where the concepts are not expected to change drastically between
the arrivals of consecutive instances. For this reason, the CVFDT algorithm of Hulten et
al. [16] checks for drift only once in a fixed number of examples (20,000). The OLIN
approach is to adjust dynamically the number of examples between model reconstructions by using the following heuristic: keep the current model for more examples
if the concept appears to be stable and reduce drastically the size of the validation
window, if a concept drift is detected.
Table 1 shows the pseudo-code outline of the OLIN algorithm. The algorithm
does some initializations and then processes a user-specified number of incoming
examples from a continuous data stream. If the user does not provide the number of the
last example to be classified by the system, the algorithm runs indefinitely. The
following parameters are calculated by OLIN: initial size of the training window, updated
window size, and the maximum difference between the training and the validation errors.
These calculations are described in the next sub-sections.
The IN algorithm, like other batch decision-tree methods, stores all the training
examples in the computer memory. Thus, applying IN to an indefinitely growing training
window is not feasible. Consequently, we have limited the maximum size of the training
window used by OLIN to Max_Win examples. This number can be adjusted to the
amount of available memory on a given computer. Another important parameter is the
training time required per each new example. As shown by us in [23], the computation
Mark Last

15

Intelligent Data Analysis

time of the IN algorithm is directly proportional to the number of training examples for
each discrete attribute and to the square of this number, if an attribute is continuous. This
imposes an additional limitation on the size of the training window, which can be handled
by a given computer system. Since the size of the training window is bounded by
Max_Win and the minimum number of examples in a validation interval is
Min_Add_Count, we can say that during the slowest periods of its operation, the training
time of OLIN per a discretely-valued new example can be proportional to Max_Win /
Min_Add_Count. With continuous attributes, this value increases to (Max_Win) 2 /
Min_Add_Count. However, when the concept appears to be stable, the algorithm training
time will go down to the order of Max_Win / Max_Add_Count or (Max_Win)2 /
Max_Add_Count, depending on the nature of the data stream attributes.

Mark Last

16

Intelligent Data Analysis

Table 1 The OLIN Algorithm

Inputs:

Sn min
n max
C
Sign
Pe
Init_Add_Count
Inc_Add_Count
Max_Add_Count
Red_Add_Count

A continuous stream of examples

The number of the first example to be classified by the system (nmin 1
examples have already arrived)
The number of the last example to be classified by the system (if
unspecified, the system will run indefinitely)
A set of candidate input attributes (discrete and continuous)
A user-specified significance level
Maximum allowable prediction error of the model
The number of new examples to be classified by the first model
Amount (percentage) to increase the number of examples between model
re-constructions
Maximum number of examples between model re-constructions
Amount (percentage) to reduce the number of examples between model reconstructions
Minimum number of examples between model re-constructions
Maximum number of examples in a training window
Info-fuzzy network

Min_Add_Count
Max_Win
IFN
Output:
Procedure OLIN
Calculate the initial size of the training window Winit (using Equation 7)
Let the training window size W = Winit
Initialize the index i of the first training example to n min - W
Initialize the index j of the last training example to W
Initialize the number of validation examples Add_Count to Init_Add_Count
While j < n max Do
Obtain a model (IFN) by applying the IN algorithm to W latest training examples
Calculate the training error rate E tr of the obtained model on W training examples
Calculate the index of the last validation example k = j + Add_Count;
Calculate the validation error rate E Val of the obtained model on Add_Count validation
examples
Update the index of the last training example j = k
Find the maximum difference between the training and the validation errors Max_Diff
(using Equation 10)
If (E Val - E tr) < Max_Diff
// concept is stable
Add_Count =Min( Add_Count * (1+
(Inc_Add_Count/100)), Max_Add_Count)
W = Min (W + Add_Count, Max_Win)
Else
//concept drift detected
Re-calculate the size of the training window W (using
Equation 8)
Update the index of the first training record i = j - W
Add_Count = Max (Add_Count * (1(Red_Add_Count/100)), Min_Add_Count)
Return the current model (IFN)
End Do

Mark Last

17

Intelligent Data Analysis

3.2. Calculating the Number of Training Examples

We assume that the OLIN system can be applied to a continuous data stream at its
beginning or after any number of incoming examples. If the data stream has just started,
the question is, how many examples need to arrive before the algorithm can induce an
initial model at a given significance level? If, on the other hand, we have an indefinite
number of past examples, what is a minimum number of latest examples that we need to
store in the computer memory? Since the IN algorithm uses an information-theoretic
heuristic in constructing the information network, the minimum number of examples
E*(z) required to confirm the statistical significance of an attribute Ai at a node z can be
found by combining Equations 2 and 3:
E* ( z)

2 (( NTi ' ( z ) 1) ( NTi ( z ) 1))

2 ln 2 MI ( Ai ' ; Ai / z )

(4)

Equation 4 is less conservative than the distribution-independent Hoeffding bound

[6], since it is based on the assumption that the split heuristic (conditional mutual
information) is distributed as chi-square with a specific number of degrees of freedom.
However, before we apply OLIN to a new data stream, we do not know the conditional
mutual information MI (Ai; Ai / z) for any of the network nodes. In fact, we do not even
know the network structure in terms of the total number of nodes and the selected
attributes to be associated with each layer. To estimate the minimum number of
examples, we first assume, without loss of generality, that the network has one layer only.
This is a good approximation, since empirical results [23] show that most nodes of the
first layer are usually terminal nodes and only a small portion of examples is associated
with subsequent layers. In general, the minimum number of examples in the training

Mark Last

18

Intelligent Data Analysis

window can always be derived from the number of examples required to split every node
in the network. We also assume that the first input attribute selected by the algorithm has
only two values: NTi (z) = 2. This number can be adjusted to the actual domain size of
candidate input attributes in a given dataset.
Since we consider only the attribute that is used to split the root node, the
conditional mutual information MI (Ai; Ai / z= 0) is equal to the mutual information MI
(Ai; Ai), which can be expressed as [5]):
MI (Ai; Ai) = H(Ai) H (Ai / Ai)

(5)

Where H(Ai) is the unconditional entropy of the target attribute Ai and H (Ai / Ai)
is the conditional entropy of Ai given an input attribute Ai. The unconditional entropy
H(Ai) can only be estimated by processing some training examples. However, it can be
approximated by its maximum value of log2 (NTi), where NTi is the number of classes [5].
If a dataset is very unbalanced, NTi may include only the most common classes. The
upper bound for the conditional entropy can be found from Fanos inequality[5]:
H (Ai / Ai) H (Pe) + Pe log2 (NTi -1)

(6)

Where Pe is the error rate of the information network having conditional entropy
of H (Ai / Ai). Since the right-hand expression is a non-decreasing function of Pe, we can
assume that Pe represents the maximum allowable error rate of the model. The
assumption of NTi (z) = 2 and Equations 4 6 lead to the following expression for the
initial size of the training window:
Winit =

2 ( NTi 1)
2 ln 2(log 2 ( NTi ) H ( Pe ) Pe log 2 ( NTi 1))

(7)

Intuitively, the expression for Winit can be interpreted as follows. A higher

significance level requires more examples, since it increases the corresponding value of
Mark Last

19

Intelligent Data Analysis

chi-square. More examples are needed to distinguish between a larger number of target
classes. Given that other factors remain unchanged, the error rate Pe is also proportional
to the number of required examples, since the algorithm needs more examples to confirm
the significance of a less accurate model. It is important to note that a high statistical
significance of the likelihood-ratio test does not necessarily imply a high predictive
accuracy of the resulting model. A significance level just represents the probability that
the model is not random, i.e., it is more accurate than the default (majority) prediction
rule.
The expression for the size of an updated window is also based on Equations 4
6, but it uses information, which is available from the latest window of the training
examples:

2 (( NTi ' 1) ( NTi 1))

W =
2 ln 2( H ( Ai ) H ( Etr ) Etr log 2 ( NTi 1))

(8)

Where NTi is the number of values (or discretized intervals) for the first attribute
Ai in the info-fuzzy network, H (Ai) is the entropy of the target, and Etr is the training
error of the current model. The resulting number of examples W should be sufficient to
confirm the significance of the first layer in the latest trained model. We assume that a
new concept, which is still unknown, can be learned from at least the same number of
examples. If a new concept pertains for fewer examples than W, the algorithm cannot
detect it at all.
3.3. Comparing the Training and the Validation Error Rates
If a concept is stable, the examples in the training window and in the subsequent
validation interval should conform to the same distribution. Consequently, there should

Mark Last

20

Intelligent Data Analysis

not be a statistically significant difference between the training and the validation error
rates of the IFN model, as we have seen in our previous studies of static datasets [23].
On the other hand, a sharp increase in the error rate may indicate a possible concept drift
[32]. Using a Normal approximation to the Binomial distribution, we calculate the
variance of the difference between error rates by the following formula (based on [27]):
Var _ Diff =

Etr (1 Etr ) Eval (1 Eval )

+
W
Add _ Count

(9)

If the concept is stable, the maximum difference between the error rates, at the
99% confidence level, is:
Max _ Diff = z0.99 Var _ Diff = 2.326 Var _ Diff

(10)

If the difference between the error rates exceeds Max_Diff, a concept drift is
detected and the size of the next training window is re-calculated by using the Equation 8.
Also, the number of examples in the next validation interval is reduced by
Red_Add_Count percent. Otherwise, the concept is considered stable and both the
training window and the validation interval are increased up to their maximum sizes.

4. Empirical Results
4.1. Manufacturing Data
We have applied OLIN to a sample of yield data recorded at a semiconductor
plant. In semiconductor industry, the yield is defined as the ratio between the number of
good parts (microelectronic chips) in a completed batch and the maximum number of
parts, which can be obtained from the same batch, if no chips are scraped at one of the
fabrication steps. Due to high complexity and variability of modern microelectronics
industry, the yield is anything but a stationary process. It is affected daily by hundreds of
Mark Last

21

Intelligent Data Analysis

material-related, equipment-related, and human-related factors. Maintaining current

models for yield prediction, as well as timely detection of changes in yield behavior,
usually called yield excursions, constitute the primary tasks of process engineers.
A semiconductor company has provided us with the records of 1,378
manufacturing batches that completed their production during a four months period. Due
to the confidentiality of the original data, we omit here the name of the company, the
description of the products, and many other details, which are irrelevant to the evaluation
of our method. The records include seven candidate input attributes, which represent the
main properties of a given batch (chip size, production priority, etc.). The target
(dependent) attribute is the percentage of outgoing yield, discretized to three intervals of
approximately equal frequency. These intervals can be characterized as a low yield, a
normal yield, and a high yield.
In our experiments, we have assumed, for the sake of convenience, that the online
learning starts after the completion of the first 378 batches, which leaves us with exactly
1,000 records for validating the performance of the constructed models. Table 2 shows
the settings used by OLIN in this dataset. The parameter values have been chosen
experimentally to provide the best classification performance of the system. All runs were
carried out on a Pentium III processor with 128 MB of RAM.

Mark Last

22

Intelligent Data Analysis

Table 2 OLIN Parameter Values (Manufacturing Data)

Parameter
n min
n max
Sign
Pe
Init_Add_Count
Inc_Add_Count
Max_Add_Count
Red_Add_Count
Min_Add_Count
Max_Win

Meaning
The number of the first example to be classified by the system
The number of the last example to be classified by the system
A user-specified significance level
Maximum allowable prediction error of the model
Initial number of examples to be classified by the first model
Amount to increase the number of examples between model reconstructions (if the model is stable)
Maximum number of examples between model re-constructions
Amount to reduce the number of examples between model reconstructions (if a concept drift is detected)
Minimum number of examples between model re-constructions
Maximum number of examples in a training window

Value
378
1377
0.1%
0.50
10
50%
100
75%
1
1,000

In Table 3, we compare the overall performance of OLIN to other methods of

online learning. The laziest approach, implemented in Run 0, is to build a model from
all the examples available at the beginning of the run (378) and to subsequently apply it
to all the validation examples (1,000) without ever re-training the IN algorithm. The no
re-training approach is based on the assumption that the data stream is stationary.
Another extreme approach (no-forgetting) is to repeatedly re-construct the
model from all the past instances after the arrival of every new example. The noforgetting learning requires longer training times as more examples arrive, and,
eventually, its training time may exceed the time between arrivals of successive examples
or surpass the limits of the computer memory. Runs no. 1 and 2 in Table 3 show the
results of this method for initial training windows of 50 and 100 examples respectively.
The next three runs (35) show the experiments, which apply static windowing by
adding and removing the same number examples from the training window at each
iteration. Finally, Run no. 6 presents the results of OLIN, which uses dynamic
Mark Last

23

Intelligent Data Analysis

windowing. The columns of the table present the total number of training windows in the
process of classifying 1,000 examples, the average number of examples in a training
window, the total run time of the system on the stream of 378 training and 1,000
validation examples, the aggregated error rate on the validation records (including its
variance), and the probability that the error rate in a given run is significantly different
from the error rate of OLIN. One and two asterisks designate 5% and 1% significance
levels respectively.
Table 3 Manufacturing Data: Summary of Experiments

Run
No.
0
1
2
3
4
5
6

Initial
Window
378
50
100
50
50
100
117

Add
Count
1000
1
1
1
10
50
Dynamic

Remove
Count
0
0
0
1
10
50
Dynamic

Number Av.
of
Window
Windows Size
1
1000
1000
550
1000
600
1000
50
100
50
20
100
23
348.2

Run
Time
(sec.)
0.14
32.57
35.81
5.16
0.60
0.28
0.66

Error
Rate
0.558
0.507
0.512
0.563
0.598
0.556
0.515

Variance
0.2466
0.2500
0.2499
0.2460
0.2404
0.2469
0.2498

p-value
0.027
0.360
0.447
0.016
0.000
0.033

*
**
*

The high error rate of the no re-training approach (Run 0) indicates that even
during a relatively short period of several months, the yield does not preserve a stationary
pattern. It is also clear that OLIN (Run 6) is significantly more accurate on this data than
the static windowing (Runs 3-5). The no-forgetting approach (Runs 1-2) appears slightly
more accurate than OLIN, but the differences are not statistically significant. In terms of
system resources, OLIN with its average window size of 348.2 examples requires more
memory than the static windowing. On the other hand, the no-forgetting method requires
an indefinite amount of memory to store its ever-growing window. In terms of run time,

Mark Last

24

Intelligent Data Analysis

OLIN is comparable with the static windowing and it is considerably faster than the noforgetting approach even for a limited stream of 1,000 records.
Figure 3 shows how OLIN adjusts the window size to the rate of concept drift. To
remove short-term effects, the error rate is averaged over past 100 validation examples.
During the first segment of the run, approximately up to Example no. 500, the error rate
keeps going up until the system recognizes a concept drift and the window size is
drastically decreased. As a result of this window adjustment, the error rate goes down
rapidly from 70% to less than 40%. However, between Examples no. 660 and 760 there
is again a sharp increase in the average error rate, followed by a very slow decline in the
error. The second peak in the error rate is not high enough to be identified as a concept
drift, and, hence, the training window continues to grow until the end of the run, where it
approaches the maximum size of 1,000 examples.
0.8

1200

0.7
1000

800

Error Rate

0.5

0.4

600

0.3
400

Number of Training Examples

0.6

0.2
200
0.1

0
350

0
550

750

950

1150

1350

Cumulative Number of Examples

Error Rate

Window Size

Figure 3 Manufacturing Data: Adjusting to Drift

Mark Last

25

Intelligent Data Analysis

Figure 4 compares the performance of four methods for online learning: no retraining, no-forgetting (initial size = 100), static windowing (size = 100, increment = 50),
and dynamic windowing (OLIN). The error rates of all methods are calculated as moving
averages of the past 100 validation examples. The no re-training approach is leading in
the beginning of the run, but eventually its error goes up and it becomes inferior to other
methods most of the time, probably due to changes in the yield behavior. The noforgetting method is consistently providing the most accurate predictions for nearly the
entire length of the data stream. The static windowing is doing better than OLIN in the
first part of the stream, up to approximately Example no. 900. Afterwards, there is a
sharp increase in the error rate of the static windowing, while OLIN and the no-forgetting
provide the lowest error. In other words, by the end of the run, the large windows of
OLIN and the no-forgetting method are more accurate than the small, fixed-size windows
of the static method.
0.8

0.7

0.6

Error Rate

0.5

0.4

0.3

0.2

0.1

0
350

550

750

950

1150

1350

Cumulative Number of Examples

OLIN

Static Window

No-forgetting

No re-training

Figure 4 Manufacturing Data: Comparison of On-line Learners

Mark Last

26

Intelligent Data Analysis

4.2. Stock Market Data

The second set of experiments has been performed on a stock market dataset,
initially used by us in [19] for evaluation of a batch learning algorithm (IFN). The raw
data represents the daily stock prices of 373 companies from the Standard & Poors 500
index [33] over a 5-year period (from 8/29/94 to 8/27/99) and it has been obtained from
the Microsoft MoneyCentral web site [34]. In [19], we have applied signal-processing
techniques to partition each series of daily stock values into a sequence of intervals
having distinct slopes (trends). An average of 15.64 intervals per company have been
identified. The classification problem has been defined as predicting the correct length of
the current interval based on the known characteristics of the current and the preceding
intervals. Consequently, we have converted every sequence of m intervals related to a
specific stock into m-1 interval-pairs each containing information about two consecutive
intervals. This resulted in a total of 5,462 records of interval-pairs. The candidate input
attributes include the duration, the slope, and the fluctuation measured in each interval, as
well as the major sector of the corresponding stock (a static attribute). The target
attribute, which is the duration of the second interval in a pair, has been discretized to
five sub-intervals of nearly equal frequency. These sub-intervals have been labeled as
very short, short, medium, etc.
To restore the original order of data arrival, we have sorted the records by the
dates of change points between the adjacent intervals. The online classification task is to
predict the timing of the next change point. To make the calculations more convenient,
we have started the online learning after the arrival of the first 462 records, which has left
us with exactly 5,000 records for validation. The parameters of OLIN were kept as

Mark Last

27

Intelligent Data Analysis

similar as possible to the settings of the experiment with semiconductor data in subsection 4.1. Due to a larger size of this dataset, the initial number of validation examples
was set to 100 and it was allowed to vary between 10 and 400 examples. The number of
the last example to be classified is 5,461. The values of other parameters remained
unchanged and the runs were performed on the same computer that was used for the first
experiment.
Table 4 compares the performance of OLIN to other methods of online learning.
Run 0 represents the no re-training learner. Runs 1 3 were performed with the noforgetting learner for a varying number of new examples accumulated before each model
re-construction. We could not reduce this number below 100 due to extremely long
training times (more than 24 hours for Add_Count = 10). Runs 4 6 show the results for
a static window size, while Run 7 represents dynamic windowing with OLIN. Like in
Table 3, one and two asterisks denote 5% and 1% significance levels respectively vs.
dynamic windowing.
Table 4 Stock Data: Summary of Experiments

Run
No.
0
1
2
3
4
5
6
7

Initial
Window
462
100
400
400
100
400
400
41

Add
Remove Number of
Count
Count
Windows
5000
0
1
100
0
50
100
0
50
200
0
25
100
100
50
100
100
50
200
200
25
Dynamic Dynamic 41

Mark Last

Av.
Window
Size
5000
2600.0
2900.0
2900.0
100
400
400
274.0

Run Time Error

(sec.)
Rate
1.26
0.450
0.384
18691.09 0.392
0.398
3.02
0.423
28.73
0.412
15.27
0.424
76.90
0.415

Variance
0.2475
0.2365
0.2383
0.2396
0.2441
0.2423
0.2442
0.2428

p-value
0.000
0.001
0.010
0.042
0.209
0.380
0.181

**
**
**
*

28

Intelligent Data Analysis

The non-stationary nature of the stock data is confirmed by the no re-training

approach having the highest error rate in Table 4 (see Run 0). The no-forgetting method
(Runs 1-3) is significantly more accurate on this data than OLIN, though the difference
between the error rates of the two methods is about 2-3% only. Table 4 also
demonstrates how fast is an increase in the computation time, if no examples are removed
from the training window: after the arrival of 5,000 records, it takes more than five hours
to get an overall accuracy, which is just by 2% higher than the accuracy, which could be
obtained within a little more than a minute! The static windowing (Runs 4-6), which
appears to be the fastest method, is not significantly worse on this dataset than OLIN,
though its error rates tend to be slightly higher. One should also note that the static
windowing provides the best result with a fixed window of 400 examples, while OLINs
dynamic windows contain an average of 274 examples only. As indicated above, the
training window size is directly related to the memory requirements of the learning
system.
Figure 5 shows how OLIN adjusts the window size to the rate of the concept drift.
It also shows the error rate averaged over the past 400 validation examples. Between the
beginning of the run and approximately Example no. 2,700, the concept appears to be
stable and there is even some decrease in the error rate vs. the initial part of the run. Due
to this apparent stability, OLIN keeps growing the window size. After Example 2700, we
observe a sharp increase in the error rate, which causes OLIN to reduce the size of the
training window. Subsequently, OLIN performs frequent updates of the window size,
trying to keep track of a rapidly changing concept. However, OLIN fails to decrease the
error rate for about the next 1,400 examples. The steep decline of the error rate in the last

Mark Last

29

Intelligent Data Analysis

500 examples, accompanied by an increase in OLINs window size, reveals the nature of
data pre-processing rather than a stock market phenomenon: the last interval of each
stock was truncated at the end of the period covered by the dataset (five years).
Consequently, the last intervals tend to be much shorter than the preceding intervals, and
their classification becomes almost deterministic.

0.7

2500

0.6
2000

1500

0.4

0.3

1000

Number of Training Examples

Error Rate

0.5

0.2
500
0.1

0
450

0
1450

2450

3450

4450

5450

Cumulative Number of Examples

Error Rate

Window Size

Figure 5 Stock Data: Adjusting to Drift

Figure 6 compares the performance of the no re-training approach, noforgetting (initial size = 100, increment = 100), static windowing (size = 400, increment =
200), and dynamic windowing (OLIN). The error rates of all methods are calculated as
moving averages of the past 400 validation examples. During the first 40% of the run,
while the concept appears to be stable, no method is consistently better or consistently
worse than the other two, which is quite reasonable. In the next segment, which is
characterized by a high rate of concept drift, the static windowing approach is less

Mark Last

30

Intelligent Data Analysis

successful than OLIN. However, OLIN itself performs worse in this segment than the
no-forgetting learner. This implies that in the second segment, retaining all the past
examples improves the classification accuracy, though it requires an increasing amount of
computer resources. A similar result has been obtained by Hulten et al. [16] in a
comparison between a static and a dynamic system. At the same time, one can see that
the gap between OLIN and the no-forgetting method remains nearly constant as more
examples arrive. It is also noteworthy that OLIN is better at adapting to a new, though
artificial concept in the last segment of the run: its error rate goes down much faster than
the error rates of the other methods, especially the no-forgetting learner. The no retraining approach is close to static windowing most of the time, but it becomes
extremely inaccurate for the last 800 examples of the run, since it does not adjust itself to
an abrupt change in the underlying concept.
0.7

0.6

Error Rate

0.5

0.4

0.3

0.2

0.1

0
500

1000

1500

2000

2500

3000

3500

4000

4500

5000

5500

Cumulative Number of Examples

OLIN

Static Window

No-forgetting

No re-training

Figure 6 Stock Data: Comparison of On-line Learners

Mark Last

31

Intelligent Data Analysis

5. Conclusions
This paper has presented OLIN, a new system for on-line classification of
continuous data streams. To sum-up, the system is based on the following principles.
The induced concept is represented in the form of an info-fuzzy network (IFN), a treelike classification model. The system is repeatedly constructing a new network from a
sliding window of latest examples. The latest model classifies the examples that arrive
before the subsequent network re-construction. A concept drift is detected by an
unexpected rise in the classification error rate. When a concept is stable, the system
increases the size of the training window, up to a pre-specified limit, and reduces the
frequency of model updates. With the detection of a concept drift, the system recalculates the size of the training window by using the principles of information theory
and statistics.
The efficiency of OLIN has been confirmed by experiments on two real-world
sets of online data. In the case of 1,000 records from a manufacturing dataset, it took
OLIN 50 times less computation time to obtain practically the same classification
accuracy like the brute force no-forgetting method. In the second experiment,
performed on 5,000 records of stock data, the ratio between the computation times of
no-forgetting and OLIN was even more dramatic: about 240. In terms of classification
accuracy, the no-forgetting approach has outperformed OLIN by 2% only.
Hulten and Domingos [15] present a number of criteria for designing an online
data mining system. These criteria include constant time and memory per incoming
record, single scan of data, availability of a usable model at any time, equivalence to an

Mark Last

32

Intelligent Data Analysis

ordinary data mining algorithm, and adaptability to time-changing concepts along with
the preservation of concepts from the past that are still relevant. With respect to these
criteria, we have seen that OLIN requires a limited amount of time and memory per
record, does not revisit old records after they are removed from the training window, uses
an anytime algorithm for constructing an info-fuzzy network (see [23]), is almost as
accurate as models induced from all the available instances and has a method for
adjusting the system parameters in response to a concept drift. The empirical results
presented in this paper show that when applied to nonstationary data, OLIN tends to be
more accurate than the static windowing methods. However, OLIN is usually less
accurate than the extremely inefficient no-forgetting approach due to apparent presence
of long-term concepts, which are eventually forgotten by OLIN, while being retained by
the no-forgetting learner.
The future work includes evaluating OLIN on more real-world datasets, as well as
on artificially built data streams. Considering the time dimension as an explicit input
attribute available to the system (similar to the approach of [3][13]) is another promising
direction. Finding a better trade-off between the efficiency of OLIN and the accuracy of
the no-forgetting method should also be examined. Other topics to study include
incremental updating of information network structure and application of a similar
approach to other classification methods. A visual environment for running beta versions
of the batch algorithm (IN) and the online algorithm (OLIN) is available for download at
http://www.ise.bgu.ac.il/faculty/mlast/ifn.htm.
Acknowledgments. This work was partially supported by the National Institute for
Systems Test and Productivity at the University of South Florida, under SPAWAR
Mark Last

33

Intelligent Data Analysis

Research Grant N00039-01-1-2248. My thanks to Prof. Eric Backer for his helpful
comments.

References
[1]

F. Attneave, Applications of Information Theory to Psychology, Holt,

Rinehart and Winston, 1959.

[2]

S.D. Bay, The UCI KDD Archive [http://kdd.ics.uci.edu], 1999.

[3]

M. Black and R. J. Hickey, Maintaining the Performance of a Learned

Classifier under Concept Drift, Intelligent Data Analysis, No. 3, pp.
453-474, 1999.

[4]

L. Breiman, J.H. Friedman, R.A. Olshen, & P.J. Stone, Classification

and Regression Trees, Wadsworth, 1984.

[5]

T. M. Cover, Elements of Information Theory, Wiley, 1991.

[6]

P. Domingos and G. Hulten, Mining High-Speed Data Streams, Proc.

of KDD 2000, pages 71-80, 2000.

[7]

W. Fan, S.J. Stolfo, J. Zhang, The Application of AdaBoost for

Distributed, Scalable and On-line Learning, Proc. of KDD-99, pages
362-366, 1999.

[8]

T. Fawcett and F. Provost, Activity Monitoring: Noticing interesting

changes in behavior, Proc. of KDD-99, 1999.

[9]

U. Fayyad, G. Piatetsky-Shapiro, and P. Smyth, From Data Mining to

Knowledge Discovery: An Overview, In U. Fayyad, G. PiatetskyShapiro, P. Smyth, and R. Uthurusamy, editors, Advances in
Knowledge Discovery and Data Mining, pages 1-30, AAAI/MIT
Press, 1996.

[10]

D. H. Fisher, Knowledge Acquisition Via Incremental Conceptual

Clustering, Machine Learning, No. 2, pp. 139-172, 1987.

[11]

M.B. Harries, C. Sammut, Extracting Hidden Context, Machine

Learning, No. 32, pp. 101-126, 1998.

[12]

D.P. Helmbold and P.M. Long, Tracking Drifting Concepts by

Minimizing Disagreements, Machine Learning, No. 14, pp. 27-45,
1994.

[13]

R. J. Hickey and M. M. Black, Refined Time Stamps for Concept Drift

Detection during Mining for Classification Rules, Proc. of
TSDM2000, LNAI 2007, pages 20-30, 2000.

Mark Last

34

Intelligent Data Analysis

[14]

C. -N. Hsu and C.A. Knoblock, Discovering Robust Knowledge from

Databases that Change, Data Mining and Knowledge Discovery, No.
2, pp. 1-28, 1998.

[15]

G. Hulten and P. Domingos, Catching Up with the Data: Research

Issues in Mining Data Streams, Proc. of Workshop on Research Issues
in Data Mining and Knowledge Discovery, 2001.

[16]

G. Hulten, L, Spencer, and P. Domingos, Mining Time-Changing Data

Streams, Proc. of KDD-2001, ACM Press, 2001.

[17]

M.G. Kelly, D.J. Hand, and N.M. Adams, The Impact of Changing
Populations on Classifier Performance, Proc. of KDD-99, pages 367371, 1999.

[18]

T. Lane and C.E. Brodley, Approaches to Online Learning and

Concept Drift for User Identification in Computer Security, Proc. of
KDD-98, pages 259-263, 1998.

[19]

M. Last, Y. Klein, A. Kandel, Knowledge Discovery in Time Series

Databases, IEEE Transactions on Systems, Man, and Cybernetics, Vol.
31, Part B, No. 1, pp. 160-169, 2001.

[20]

M. Last, A. Kandel, O. Maimon, Information-Theoretic Algorithm for

Feature Selection, Pattern Recognition Letters, Vol. 22, No. 6, pp. 799811, 2001.

[21]

N. Littlestone, Learning Quickly When Irrelevant Attributes Abound:

A New Linear-threshold Algorithm, Machine Learning, No. 2, pp.
285-318, 1988.

[22]

H. Liu and H. Motoda, Feature Selection for Knowledge Discovery

and Data Mining, Kluwer, 1998.

[23]

O. Maimon and M. Last, Knowledge Discovery and Data Mining, The

Info-Fuzzy Network (IFN) Methodology, Kluwer Academic
Publishers, 2000.

[24]

O. Maimon, A. Kandel, and M. Last, Information-Theoretic Fuzzy

Approach to Data Reliability and Data Mining, Fuzzy Sets and
Systems, Vol. 117, No. 2, pp. 183-194, 2001.

[25]

U. Manber, A. Patel, and J. Robison, Experience with Personalization

on Yahoo!, Communications of the ACM, Vol. 43, No. 8, pp. 35-39,
2000.

[26]

W. Mendenhall, J.E. Reinmuth, R.J. Beaver, Statistics for

Management and Economics, Duxbury Press, 1993.

[27]

T.M. Mitchell, Machine Learning, McGraw-Hill, 1997.

Mark Last

35

Intelligent Data Analysis

[28]

J.R. Quinlan, Induction of Decision Trees, Machine Learning, Vol. 1,

No. 1, pp. 81-106, 1986.

[29]

J. R. Quinlan, C4.5: Programs for Machine Learning, Morgan

Kaufmann, 1993.

[30]

C.R. Rao and H. Toutenburg, Linear Models: Least Squares and

Alternatives, Springer-Verlag, 1995.

[31]

P.E. Utgoff, An Improved Algorithm for Incremental Induction of

Decision Trees, Proc. of the Eleventh International Conference on
Machine Learning, pages 318-325, 1994.

[32]

G. Widmer and M. Kubat, Learning in the Presence of Concept Drift

and Hidden Contexts, Machine Learning, Vol. 23, No. 1, pp. 69-101,
1996.

[33]

Standard & Poors Index at http://www.spglobal.com/

[34]

The Microsoft MoneyCentral home page at

http://moneycentral.msn.com/

Mark Last

36