Use of ensembles of Fourier spectra in capturing recurrent concepts in data streams

Full text

Use of Ensembles of Fourier Spectra in Capturing
Recurrent Concepts in Data Streams
Sripirakas Sakthithasan and Russel Pears
School of Computer and Mathematical Sciences
Auckland University of Technology, New Zealand
Email: {ssakthit, rpears }@aut.ac.nz
Albert Bifet and Bernhard Pfahringer
Department of Computer Science
University of Waikato, New Zealand
Email: {albifet , bfahringer }@cs.waikato.ac.nz
Abstract—In this research, we apply ensembles of Fourier
encoded spectra to capture and mine recurring concepts in a data
stream environment. Previous research showed that compact ver-
sions of Decision Trees can be obtained by applying the Discrete
Fourier Transform to accurately capture recurrent concepts in a
data stream. However, in highly volatile environments where new
concepts emerge often, the approach of encoding each concept
in a separate spectrum is no longer viable due to memory
overload and thus in this research we present an ensemble
approach that addresses this problem. Our empirical results on
real world data and synthetic data exhibiting varying degrees
of recurrence reveal that the ensemble approach outperforms
the single spectrum approach in terms of classification accuracy,
memory and execution time.
I. INTRODUCTION
In many real world applications, patterns or concepts recur
over time. Machine learning applications that model, capture
and recognize concept re-occurrence gain significant efficiency
and accuracy advantages over systems that simply re-learn
concepts each time they re-occur. When such applications
include safety and time critical requirements, the need for
concept re-use to support decision making becomes even more
compelling.
Auto-pilot systems sense environmental changes and take
appropriate action (classifiers, in the supervised machine learn-
ing context) to avoid disasters and to fly smoothly. As en-
vironmental conditions change, appropriate actions must be
taken in the shortest possible time in the interest of safety.
Thus for example, a situation that involves the occurrence
of a sudden low pressure area coupled with high winds (a
concept that would be captured by a classifier) would require
appropriate action to keep the aircraft on a steady trajectory. A
machine learning system that is coupled to a flight simulator
can learn such concepts in the form of classifiers and store
them in a repository for timely re-use when the aircraft is on
live flying missions. In live flying mode the autopilot system
can quickly re-use the stored classifiers when such situations
re-occur. Additionally, in live flying mode, new potentially
hazardous situations not experienced in simulator mode can
also be learned and stored as classifiers in the repository for
future use.
In a real world setting, there is an abundance of applications
that exhibit such recurring behavior such as stock and sales
applications where timely decision making results in improved
productivity. Our research setting is a data stream environment
where we seek to capture concepts as they occur, store them
in highly compressed form in a repository and to re-use
such concepts for classification when the need arises in the
future. A number of challenges need to be overcome. Firstly,
a compression scheme that captures concepts using minimal
storage is required as in a high volatile high dimensional
environment. Memory overhead will be a prime concern as
the number of concepts will grow continuously in time given
the unbounded nature of data streams. Secondly, in real-world
environments, concepts rarely, if ever, occur in exactly their
original form and so a mechanism is needed to recognize par-
tial re-occurrence of concepts. Thirdly, the concept encoding
scheme needs to be efficient in order to support high speed
data stream environments.
In order to meet the above challenges, we extend the work
proposed in [16] in a number of ways. In [16] concepts were
initially captured using decision trees and the Discrete Fourier
Transform (DFT) was applied to encode them into spectra
yielding compressed versions of the original decision trees.
Firstly, instead of encoding each concept using its own
Fourier spectrum, we use an ensemble approach to aggregate
individual spectra into a single unified spectrum. This has two
advantages, the first of which is reduction of memory overhead.
Memory is further reduced as Fourier coefficients that are
common between different spectra can be combined into a
single coefficient, thus eliminating redundancy. The second
advantage arises from the use of an ensemble: new concepts
that manifest as a combination of previously occurring con-
cepts already present in the ensemble have a higher likelihood
of being recognized, resulting in better accuracy and stability
over large segments of the data stream.
Secondly, we devise an efficient scheme for spectral energy
thresholding that directly controls the degree of compression
that can be obtained in encoding concepts in the repository.
Thirdly, we optimize the DFT encoding process by re-
moving the need for computing a potentially expensive inner
product operation on vectors.
II. RE LATE D RESEARCH
While a vast literature on concept drift detection exists
[13], only a small body of work exists so far on exploitation
of recurrent concepts. The methods that exist fall into two
broad categories. Firstly, methods that store past concepts as
models and then use a meta-learning mechanism to find the
best match when a concept drift is triggered [5], [7]. Secondly,
methods that store past concepts as an ensemble of classifiers.

The method proposed in this research belongs to the second
category where ensembles remember past concepts.
An algorithm called REDDLA is presented in [14]. This
algorithm is designed to handle recurring concepts with un-
labeled data instances. One of the key issues is that explicit
domain is required on the concept recurrence interval. The
other issue is high memory overhead.
Lazarescu in [11] proposed an evidence forgetting mecha-
nism based on a multiple window approach and a prediction
module to adapt classifiers based on estimating future rate
of change. Whenever the difference between observed and
estimated rates of change are above a threshold, a classifier
that best represents the current concept is stored in a reposi-
tory. Experimentation on the STAGGER data set showed that
the proposed approach outperformed the FLORA method on
classification accuracy with re-emergence of previous concepts
in the stream.
Ramamurthy and Bhatnagar [15] use an ensemble approach
based on a set of classifiers in a global set G. An ensemble of
classifiers is built dynamically from a collection of classifiers
in G, if none of the existing individual classifiers are able to
meet a minimum accuracy threshold based on a user defined
acceptance factor. Whenever the ensemble accuracy falls below
the accuracy threshold, G is updated with a new classifier
trained on the current chunk of data.
Another ensemble based approach by Katakis et al. is
proposed in [9]. A mapping function is applied on data stream
instances to form conceptual vectors which are then grouped
together into a set of clusters. A classifier is incrementally built
on each cluster and an ensemble is formed based on the set
of classifiers. Experimentation on the Usenet data set showed
that the ensemble approach produced better accuracy than a
simple incremental version of the Naive Bayes classifier.
Gomes et al. [7] used a two layer approach with the first
layer consisting of a set of classifiers trained on the current
concept, while the second layer contains classifiers created
from past concepts. A concept drift detector flags when a
warning state is triggered and incoming data instances are
buffered to prepare a new classifier. If the number of instances
in the warning window is below a threshold, the classifier in
layer 1 is used instead of re-using classifiers in layer 2. One
major issue with this method is validity of the assumption that
explicit contextual information is available in the data stream.
Gama and Kosina also proposed a two layered system in
[5] which is designed for delayed labelling, similar in some
respects to the Gomes et al. [7] approach. In their approach,
Gama and Kosina pair a base classifier in the first layer
with a referee in the second layer. Referees learn regions of
feature space which its corresponding base classifier predicts
accurately and is thus able to express a level of confidence on
its base classifier with respect to a newly generated concept.
The base classifier which receives the highest confidence score
is selected, provided that it is above a user defined hit ratio
parameter; if not, a new classifier is learnt.
Just-in-Time classifiers is the solution proposed by Allipi
et al. [1] to deal with recurrent concepts. Concept change
detection is carried out on the classification accuracy as well as
by observing the distribution of input instances. The drawback
is that this model is designed for abrupt drifts and is weak at
handling gradual changes.
Recently, Sakthithasan and Pears in [16] used the Discrete
Fourier Transform (DFT) to encode decision trees into a
highly compressed form for future use. They showed that DFT
encoding is very effective in improving classification accuracy,
memory usage and processing time in general. It maintains a
pool of Fourier spectra and a decision tree forest in parallel.
The decision tree forest dominates the model, when none of the
existing Fourier spectra matches the current concept, otherwise
classification is done by the best performing Fourier spectrum.
III. APP LI CATI ON O F TH E DISCRETE FOURIER
TRANSFORM ON DECISION TRE ES
The Discrete Fourier Transform (DFT) has a vast area of
application in diverse domains such as time series analysis,
signal processing, image processing and so on. It turns out
as Park [12] and Kargupta [10] show, that the DFT is very
effective in terms of classification when applied on a decision
tree model.
Kargupta et al. [10], working in the domain of distributed
data mining, showed that the Fourier spectrum fully captures
a decision tree in algebraic form, meaning that the Fourier
representation preserves the same classification power as the
original decision tree.
A. Transforming Decision Tree into Fourier Spectrum
A decision tree can be represented in compact algebraic
form by applying the DFT to paths of the tree. Each Fourier
coefficient ωjis given by:
ωj=1
2dX
x
f(x)ψλ
j(x); (1)
ψλ
j(x) = Qmexp 2πi
λmxmjmwhere j and x are strings of length
d;xmand jmrepresent the mth attribute value in jand
xrespectively; f(x)is the classification outcome of path vector
x and ψλ
j(x)is the Fourier basis function.
Fig. 1. Decision Tree with 3 binary features
Figure 1 shows a simple example with 3 binary valued
features x1,x2and x3, out of which only x1and x3are
actually used in the classification.
With the wild card operator * in place we can use equations
(1) and (2) to calculate non zero coefficients. Thus for example
we can compute:
ω000 =4
8f(∗ ∗ 0)ψ000(∗ ∗ 0) + 2
8f(0 ∗1)ψ000(0 ∗1)
+2
8f(1 ∗1)ψ000f(1 ∗1) = 3
4
ω001 =4
8f(∗ ∗ 0)ψ001(∗ ∗ 0) + 2
8f(0 ∗1)ψ001(0 ∗1)
+2
8f(1 ∗1)ψ001f(1 ∗1) = 1
4

and so on.
Kargupta et al in [10] showed that the Fourier spectrum of
a given decision tree can be approximated by computing only a
small number of low order coefficients, thus reducing storage
overhead. With a suitable thresholding scheme in place, the
Fourier spectrum consisting of the set of low order coefficients
is thus an ideal mechanism for capturing past concepts.
Furthermore, classification of unlabeled data instances can
be done directly in the Fourier domain as it is well known that
the inverse of the DFT defined in expression 1 can be used
to recover the classification value, thus avoiding the need for
expensive reconstruction of a decision tree from its Fourier
spectrum. The inverse Fourier Transform is given by
f(x) = X
j
ωjψλ
j(x)(2)
where ψλ
j(x)is the complex conjugate of ψλ
j(x). ‘ An instance
can be transformed into binary vector through the symbolic
mapping between the actual attribute value and mapped value
( either 0 or 1 in binary case). It can then be classified using
the inverse function in equation 2. Suppose the instance is 010,
the classification value f(010) can be calculated as follows:
f(010) = 1
2d(−1)000.010ω000 +1
2d(−1)001.010ω001
+1
2d(−1)010.010ω010 +1
2d(−1)011.010ω011
+1
2d(−1)100.010ω100 +1
2d(−1)101.010ω101
+1
2d(−1)110.010ω110 +1
2d(−1)111.010ω111 = 1 (3)
IV. EXP LO ITATI ON O F TH E FOURIER TRANSFORM FOR
RECURRENT CON CE PT CA PT UR E
We first present the basic algorithm used in Section IV-A
and then go on to discuss an optimization that we used for
energy thresholding in Section IV-B.
We use CBDT [8] as the base classifier which maintains a
forest of Hoeffding Trees [4] CBDT is dynamic in the sense
that it can adapt to changing concepts at drift detection points.
As shown in Figure 2, the memory is divided into two
segments: the forest of Hoeffding trees; and a pool of Fourier
Spectra. The forest learns and undergoes structural modifica-
tion on a continuous basis. The pool maintains a collection
of Fourier Spectra encoded from Hoeffding Trees, each of
which had the best classification accuracy across the forest at a
particular concept drift point. Each Hoeffding Tree and Fourier
Spectrum is equipped with an instance of a drift detector. In
this research, we use the SeqDrift2 drift detector [13] as the
default option.
In [16], each Fourier spectrum is represented individually
as a Fourier Concept Tree (FCT). In this work, we aggregate
spectra and maintain a pool of ensemble spectra known as
Ensemble Pool (EP). The aggregation process is carried out in
two different ways. Algorithm EPaaggregates with reference
to similarity based on accuracy whereas EP aggregates based
Fig. 2. An Architecture for Recurrent Concept Capture
on structural similarity. We describe the EP process in Algo-
rithm EP and discuss how FCT can be generated from it as a
special case.
In practice, any incremental decision tree approach that
uses a forest of trees can be used in place of CBDT base
classifier.
A. EP Algorithm
Algorithm EP
Input: Energy Threshold ET, Accuracy Tie Threshold τ
Input: Structural Similarity Threshold α
Output: Best Performing Classifier Cthat suits current concept
1. Plant a Hoeffding tree rooted on each attribute found in the data
stream
2. Cis set to a randomly selected Hoeffding tree model from forest
3. Initialise an empty pool
4. Read an instance Ifrom the data stream
5. repeat
6. Apply all classifiers in forest and pool to classify I
7. Append 0to the embedded drift detector’s window for each
classifier if classification is correct, else 1
8. until Drift is detected by the current best classifier C
9. if C is from the forest
10. Identify best performing Fourier Spectrum Fin pool
11. if (accuracy(C)-accuracy(F))> τ
12. Apply DFT on model Cto produce Fourier
Spectrum F* using energy threshold ET
13. if F* is not already in pool
14. Call Aggregation
15. Identify best classifier Cacross forest and pool
16. GoTo 4
Algorithm Aggregation
Input: Fourier Spectrum F*,a set of existing ensembles Ein pool
Output: Updated Pool
1. repeat Over all data instances i
2. for Each ensemble Ein the pool
3. d(E) = d(E) + |c(F*, i)−c(E, i)|
4. until Next concept drift point
5. E∗= arg min
E
(distance(E))
6. if (E* ≥α) Merge F* with E
7. else insert E as a new spectrum
In step 1, a Hoeffding Tree rooted on each attribute is

created. In step 2, a tree is randomly chosen as the best
performing classifier C. Next, an empty pool is created in step
3. Each incoming instance is routed to all trees in the forest and
pool until a concept drift signal is triggered by the drift detector
instance attached to the best classifier C(steps 4 to 8). At the
first concept drift point, the best performing tree C(in terms
of drift detector’s estimate of accuracy) is transformed into
a Fourier Spectrum F* after energy thresholding [16]. In this
method, the assumption is that the best tree that has the highest
accuracy helps locate conceot changes precisely than other
trees because it is the tree that captures concepts at a greater
detail than others, thus the highest accuracy. Thereafter, F* is
stored in the repository for reuse whenever the concept recurs.
The spectra stored in the repository are fixed in nature as the
intention is to capture past concepts. A new best performing
classifier is then identified as shown in step 15.
At each subsequent drift point, if the best classifier is from
the pool then that classifier is applied to classify data instances
until a new best classifier emerges at a subsequent drift point.
Otherwise, if the best classifier is from forest, two tests are
made prior to applying the DFT to reduce redundancy in the
pool. Firstly, we check whether the difference in accuracy
between the best Hoeffding tree in forest (C) and the best
performing Fourier Spectrum (F) in the pool (from step 10) is
greater than a user defined tie threshold τ(step 11). If this test
succeeds, the DFT is applied to C to produce (F*) (step 12).
Furthermore, a second test is made to ensure that its Fourier
representation (F*) is not already in the pool (step 13). If this
test is also passed, algorithm Aggregation is called to integrate
F* into a selected existing Fourier Spectrum (E*) or plant (F*)
as a separate Fourier spectrum in the pool (step 14).
Algorithm Aggregation searches for the spectrum (E*)
that has the greatest structural similarity to the currently
generated spectrum (F*) (step 3). Step 3 evaluates the degree
of disagreement (d) between the classification decisions (c) for
F* and E on data instance i. Degree of disagreement between
(F*) and each of the existing ensemble (E) in pool can easily be
updated incrementally in Algorithm EP using a single counter
variable at each ensemble E. This removes the steps from 2 to
4 in Algorithm Aggregation. As an alternative to aggregating
structurally similar spectra, we used accuracy as the measure
that defines similarity. Similarity based on accuracy leads to
aggregating similar performing Fourier Spectra together. Thus,
we test the hypothesis, aggregation of two spectra based on
structural similarity produces better performing trees than the
one based on accuracy.
As stated earlier, FCT omits the call to Algorithm Aggre-
gation and inserts (F*) as it is, and is thus a special case of
EP.
B. Optimising the Energy Thresholding Process
Sakthihasan et al. in [16] showed that classification ac-
curacy is sensitive to spectral energy, which is given by the
total of the sum of squares of the coefficients[10]); the higher
the energy the greater is the classification accuracy in general.
Thresholding on spectral energy is thus an effective method of
obtaining a compact spectrum while retaining the classification
power inherent in the decision tree counterpart.
A solution described in [16] was to iterate through each
order of the spectrum and compute ratio of energy at orders
i−1to that of irespectively. Thresholding can then be
implemented at order O when the ratio is less than some
small tolerance value, say 0.01. The drawback of this simple
solution is that it does not guarantee that the cumulative
energy up to order O contains a proportion () of the
total energy. Fortunately, a solution exists for this problem.
Theorem 1 proves that E(T) (total energy of Fourier Spectrum)
equals to ω0(The 0th coefficient). Thus, total energy can
be computed efficiently, without having to enumerate all the
single coefficients.
Theorem 1 The total spectral energy E=Pjω2
j=ω0,
where ω0denotes the coefficient with order 0, which is easily
computed as its Fourier basis function is unity.
Proof: Omitted due to lack of space and can be found
in http://cogprints.org/9879/
This optimization significantly increases processing speed,
especially in high dimensional data stream environments.
The next optimization is applied to optimize the Fourier
Basis function calculation in equation 1 especially when
wildcard characters (denoting absence of a feature) are
present in a path vector xof a Hoeffding Tree.
C. Optimizing the Computation of the Fourier Basis Function
The computation of a Fourier basis function for a given
partition jin a generic n−ary (n≥2) domain is given by:
X
x∈S
ψj(x) = X
x∈SY
m
exp2πijmxm
λm(4)
Thus we can see from (4.3) that the computation of
Px∈Sψ(j)over a set of schema S requires the computation
of an expensive inner product operation between the xand
j. However, it is possible to optimize this inner product
computation as defined in Theorem 2.
Theorem 2 The computation of Px∈Sψj(x)can be
optimized as follows:
Case 1: If there exists at least one (p, ∗)combination
with p∈j,p6= 0 and ∗a wild card character defining a set
of schema S, then Px∈Sψj(x)=0.
Case 2: else if there exists ncombinations of (0,∗)
pairs in the jand xvectors respectively, then
X
x∈S
ψj(x) = λ
λk−1
Y
k=n
exp2πijkxk
λkwhere λ=Qn−1
l=0 λl
Proof: Omitted due to lack of space and can be accessed
from:http://cogprints.org/9879/
The value of Case 1 is that a simple scan of the jand
xvectors will save a total of dmultiplications and d−1
additions.

We now turn our attention to Case 2. Since Qlλlis a
constant for all possible values of jand y, the value of Case
2 is that a scan of the two vectors will avoid the overhead of
nmultiplications and n−1additions.
Even with these optimizations, coefficient calculation may
be expensive in a large dimensional data set. In the next section
we present a strategy to further optimize the derivation of the
spectrum.
D. Localized Approach to Ensemble Learning in the Fourier
Domain
In order to realize the full benefits of ensemble learning in
the Fourier domain, we aggregate individual spectra si(x)that
represent different concepts which manifest at different points
in the stream.
sc(x) = X
i
AiX
i
si(x)
=X
i
AiX
j∈Pi
ωj(i)ψj(x)(5)
where sc(x)denotes the ensemble spectrum produced from the
individual spectra si(x)produced at different points iin the
stream; Aiis the classification accuracy of its corresponding
spectrum and Piis the set of partitions for non zero coefficients
in spectrum si.
Park in [12] used ensemble learning with Fourier spectra
in a setting different to ours. They considered a distributed
system with each node iproducing its own spectrum si(x)
and aggregation taking place at a central node. In our setting
of a data stream environment, we do not have all spectra in
advance but we can still use the same principle due to the
distributive nature of the linear weighted sum expressed by
(5). Hence, we use:
s(i+1)
c(x) = s(i)
c(x) + Ai+1si+1 (x)(6)
where s(i+1)
c(x),s(i)
crepresent the ensemble spectra at concept
drift points i+ 1 and irespectively in the stream and si+1(x)
is the spectrum produced at drift point i+ 1 with accuracy
Ai+1.
We use expression (6) for implementing ensemble learning
but with one essential difference. A direct application of (6)
using the entire (global) set of attributes Gcomprising the data
set would be inefficient. As there are an exponential number
of coefficients with respect to the number of attributes, this
could cause a bottleneck in high dimensional environments.
One practical solution is to populate the spectrum using only
attributes present in a given tree. The major advantage of this
approach is smaller computational overhead as the Fourier
transform effort is directly proportional to the size of the
attribute set used. Then this initial spectrum can be extended to
a full length spectrum containing the attributes that are absent
in the given tree, using a simple transformation scheme.
We define an attribute set of a Decision Tree as that subset
of attributes which define splits in the tree. Suppose that we
are integrating spectra from trees D1and D2, having attribute
sets Land Mrespectively. We apply the DFT on D1to obtain
S1using only the attributes in its attribute set Land not all
attributes in G. Similarly we generate S2from D2using only
the attributes defined in M.
Now, in order to integrate S1with S2, we need to account
for differences in the attribute sets Land M. To do this, we
take S1and expand the spectrum by incorporating attributes in
the set M\L. The expansion is defined by a single operation:
For each schema instance in the spectrum (say S1) expand
the spectrum by adding 0to all attribute index positions in
set M\L. The coefficient value after expansion will remain
it the same as the classification fvalue for all of these added
index positions remains unchanged. We are now in a position
to integrate two spectra produced from their own localized
set of attributes. Essentially, this means that we now have
a more efficient method of implementing ensemble learning
using expression (6).
The next section presents the empirical outcomes of the
proposed models with the above mentioned optimizations.
V. EXPERIMENTAL STU DY
The main focus of the study is to assess the effectiveness
of the ensemble EP approach vis-a-vis FCT in respect of
classification accuracy, memory consumption, processing
speed, tolerance to noise. We also assessed the sensitivity
of EP’s accuracy on two significant factors, pool size and
impact of drift detector. All experimentation was done with
the following parameter values:
Tree Forest: Max Node Count=5000, Max Number of
Fourier spectra=10, Tie Threshold τ=0.01
SeqDrift2/ADWIN [2]: drift significance value=0.01
A. Datasets Used for the Experimental Study
1) Synthetic Data: We experimented with the Rotating
Hyperplane data generator that is commonly used in drift
detection and recurrent concept mining. The dataset was
generated within the MOA data stream tool [3]. We injected
concept recurrence into the stream at known points so that we
could evaluate the capabilities of FCT and EP to recognize and
exploit such recurrences. For this dataset 10 different concepts
were generated, each of which spanned 5,000 instances and
each occurred a total of 3 times at different points in the
stream. In order to challenge the concept recognition process,
we added 10% noise by inverting the class labels of 10% of
randomly selected instances.
2) Real World Data: Spam Data Set: The Spam dataset
was used in its original form 1which encapsulates an evo-
lution of Spam messages. There are 9,324 instances and 499
informative attributes.
Electricity Data Set: NSW Electricity dataset is also used
in its original form 2. There are two classes Up and Down
that indicate the change of price with respect to the moving
average of the prices in last 24 hours.
Flight Data Set: This dataset is generated through the
use of NASA’s FLTz flight simulator which was designed to
simulate flight conditions experienced with commercial flights.
1from http://www.liaad.up.pt/kdus/products/datasets-for-concept-drift
2from http://moa.cms.waikato.ac.nz/datasets/

It consists of a set of 20 separate files, each containing data
about a single flight with four scenarios: take off, climb,
cruise and landing. Data is recorded every second and a data
instance is produced. The ”Velocity” feature is chosen as the
class feature as it needs to be adjusted in order to maintain
aircraft stability during various maneuvers such as take off
and landing. Velocity was discretized into binary outcomes
”UP” or ”DOWN” depending on the directional change of the
moving average in a window of size 10 data instances.
B. Comparative Study: Ensemble versus Single Spectrum Ap-
proach
Previous research on the use of Fourier spectrum re-
vealed accuracy and memory advantages over meta learning
approaches such as the one employed by Gama and Kosina
in storing past concepts in a repository [16]. For details of
the advantages of the Fourier approach and experimentation
with it the reader is referred to [16]. Our focus here is a
comparative study of the Ensemble approach versus the single
spectrum approach. With this in mind we designed three types
of experiments.
1) Accuracy: Accuracy is a critical performance measure
in many practical applications. Due to the dynamic nature of
data streams classification accuracy on the current concept was
taken as the performance measure. Figure 3 presents Accuracy
Fig. 3. Accuracy Profiles
values of all algorithms at 10 equal-sized sub-divisions of the
stream. We also present overall mean and standard deviations
of accuracy taken across the entire stream for each dataset.
Fig 3 shows that the individual accuracies across segments
and overall accuracy across the entire stream are consistent
with each other. MetaCT, which uses a referee based strategy
was found to be the worst performing algorithm on all datasets.
In contrast EP outperforms the other algorithms in general,
followed by EPaand FCT. These results show clearly that
DFT based methods are superior in a dynamic data stream
environment.
FCT does not exploit aggregation of Fourier Spectra and
is hence challenged in a memory constrained environment
where the number of models stored for reuse is limited. Figure
3 depicts the performance in such an environment where
memory is severely limited. This introduces a large burden
on FCT to re-learn concepts after change. EP is more resilient
at small pool sizes as any given concept that recurs can be
approximated by a linear combination of spectra embedded
in the ensemble, just as a waveform of arbitrary shape can
be approximated by a large enough sum of sine functions in
signal processing.
Examining model usage statistics, EP was 3.7 times higher
in model re-use on the Flight dataset. The corresponding value
was 2.6 for EPaon the same dataset. This provides empirical
support for the claim that an aggregation-based model such
as EP has a significant advantage in reducing the degree of
relearning. For Rotating Hyperplane with known recurrence
points the advantage of EP over its counterparts is very explicit.
We display the stream segment for the third round of concept
occurrences, spanning the 10 concepts. Each of the 10 intervals
represent the second recurrence of a concept and the Figure
shows that EP outperforms FCT on 8/10 concepts; EPaand
MetaCT on 7/10 concepts; and CBDT on all 10 concepts.
The next key aspect in a memory constrained environment
is memory consumption which is assessed in the following
section.
2) Memory: Memory consumption is influenced by the
degree of generalizability of a given algorithm. A greater
degree of generalizability promotes higher re-use and reduces
the number of spectra that need to be stored in the repository
to achieve a given level of classification accuracy. In this
context it will be interesting to compare the consumption of
EP with that of FCT as they have contrasting model re-use
characteristics.
MetaCT(SeqDrift2) and CBDT were excluded from mem-
ory comparison due to their relatively poor performance in the
previous experiment.
TABLE I. MEMORY USAGE WITH POOL S IZE S ET T O 10
Dataset Average Pool Memory (in KBs)
FCT EPaEP
Flight 32.1 20.2 18.1
Electricity 31.6 16.1 14.1
Rot. Hyperplane 48.4 38.6 27.9
Spam 17.3 17.2 16.4
Table I presents the average memory consumption of the pool
over the entirety of each dataset. As mentioned in Section III,
each of the above algorithms in Table I has two components:
a forest and a repository pool. Memory consumed by forest
is not a distinguishing factor as there was a very marginal
difference between the algorithms and thus the focus was on
the repository pool.

Without exception, EP consumed the least memory com-
pared to the other algorithms. This was expected as EP
structurally examines instance vectors (i.e. corresponding to
classification paths in Hoeffding tree) and aggregates similar
vectors together. On the other hand in EPa, structural sim-
ilarity is not guaranteed and two structurally very different
Spectra producing similar accuracy could be chosen as the
candidates to be aggregated, thus resulting in larger spectra.
Table I provides evidence to support this premise as the
memory consumed by EPais higher than that of EP but
lower than FCT. On average over all datasets, EP achieved a
41% reduction in memory consumption in relation to FCT; the
corresponding figure for Electricity was 55%. This represents
a significant benefit of applying aggregation in Fourier space.
3) Processing Speed: DFT application is a potential perfor-
mance bottleneck when compared to classification, especially
in high dimensional data streams.
Processing speed is dependent on a variety of factors:
maintaining and classifying relatively larger number of Fourier
Spectra in FCT compared to EP and EPa, aggregation in EP
and EPathat generalize models thus reducing re-learning and
the need for DFT application, and finally the computational
overheads of aggregation. Therefore, this section assesses the
trade off between single and aggregated Fourier approaches in
terms of processing speed.
TABLE II. PRO CES SI NG SPE ED I N INS TANC ES P ER SE CO ND
Dataset FCT EPaEP
Flight 797.2 731.2 836.9
Electricity 11600.3 9002.5 11402.5
Rotating Hyperplane 5647.8 5413.8 5804.5
Spam 4.2 3.9 4.2
Table II shows that EP is the fastest most of the time. EPa,
even though it has the potential to be faster due to its simple
aggregation strategy, suffers from inappropriate aggregations
that introduce instability, thus triggering more drift points than
its EP counterpart. EP, on the other hand, efficiently does
structural similarity comparison by incrementally updating
simple counters that remembers the number of disagreements
in classification between the current winner tree and every
Fourier Spectra in pool. On the other hand, although EP
through its aggregation strategy requires more computational
effort than EPa, that effort is compensated for by its stability,
which triggers fewer false drift alarms than either EPaor FCT.
Therefore, this experiment demonstrates that an expensive
operation such as aggregation if applied appropriately will
yield a direct processing speed advantage over a period of
time.
4) Effects of Noise: Algorithms that work well in noise-
free environments will fail on noisy environments if they
lack the ability to generalize to new data by removing minor
variations which often correspond to noise. DFT application, as
mentioned earlier, extracts significant coefficients by ignoring
minor coefficients that may capture noise inherent in data. It
was shown in [16] that DFT application provides robustness in
a noisy environment as opposed to a non-DFT based approach
such as MetaCT. Therefore, this experiment is aimed at testing
whether aggregation has an added advantage over a non-
aggregation based method such as FCT.
Fig. 4. The impact of noise on accuracy
Figure 4 shows percentage accuracy decrease for noise
levels 20% and 30% on FCT and EP relative to accuracy on the
original Flight dataset. It is clear that the decrease in accuracy
is higher at the 30% noise level. What is interesting is the
higher tolerance of EP to noise compared to FCT. In 8/10
intervals, for 20% noise, EP is found to be having a lesser
decrease than its counterpart. Similarly at the 30% noise level,
the fraction is 4/10, with the two being tied in performance in
two other intervals. Again, as with the other metrics that we
tracked, the superior performance of EP can be explained in
terms of its power to generalize making it more robust to the
effects of noise 3.
Next we examine the sensitivity of EP on key parameters
that significantly affect performance. Due to the superiority of
EP over the other algorithms, the study was confined to this
algorithm. Please refer [16] for sensitivity analysis on FCT’s
parameters.
C. Sensitivity Analysis
EP(SeqDrift2) has two key parameters of its own: pool size
and choice of drift detector.
1) Pool Size: In this experiment we contrasted classifica-
tion accuracy at two different ends of the pool size scale,
namely 1 and 10. In the context of the Flight dataset which
has four concepts, a pool size of 1 represents an extremely
limiting memory environment and the size of 10 represents a
situation where memory is plentiful. Figure 5 shows accuracy
Fig. 5. The impact of pool size on flight dataset
values over 10 intervals. Interestingly, EP, with pool size1,
has the highest accuracy in 8/10 intervals. There is a 7.6%
and 7.2% gain in accuracy compared to FCT over pool sizes
1 and 10 respectively. This is a significant outcome of this
research. Even in an extreme memory challenged environment,
EP achieves its best accuracy over a setting with a much
3The other 3 datasets that we experimented with displayed similar trends
to that of the Flight dataset and were thus not included in interests of space
constraints

higher memory capacity. The implication is that ensemble
accuracy increases with greater diversity and resonates with
the research conducted by [6]. This illustrates the strength of
aggregation applied in the EP algorithm. As more memory
becomes available at pool size 10, FCT’s accuracy converges
to that of its counterpart, as expected. At the higher memory
setting FCT can accommodate more spectra in its pool that
are tailored to specific concepts.
2) Impact of Drift Detector: A drift detector that incor-
rectly triggers change points leads to partial learning of a con-
cept and under developed classifiers being stored in the pool.
This introduces fluctuations in accuracy, which in turn trigger
change detections, causing even more fluctuations and so on.
This is a cyclic problem. On the other hand, if a drift detector
fails to detect changes, classifiers are not updated in a timely
fashion, thus leading to poor performance. This situation may
arise if a drift detector has significantly high detection delay in
signaling changes. The ADWIN and SeqDrift2 drift detectors,
as shown in [13] have contrasting properties. SeqDrift2 has a
lower false positive rate than ADWIN while having similar
sensitivity to ADWIN. Therefore, the comparative study is
largely governed by false positive detections.
Fig. 6. The impact of drift detector on EP with pool size 10
Figure 6 reveals that SeqDrift2 helped EP to reduce the
frequency of sudden accuracy drops seen with ADWIN, due
to the latter signaling false changes in concepts. In the segment
shown in Figure 6, there is a 5% gain in accuracy by using
SeqDrift2 and it is 3.4% over the entire data set.
VI. CONCLUSIONS AND FUTURE WO RK
In this research we proposed a novel approach for cap-
turing and exploiting recurring concepts in data streams. We
optimized the derivation of the Fourier spectrum by employing
two mechanisms: one for energy thresholding and the other for
speeding up computation of the Fourier basis functions.
This research revealed that the ensemble approach outper-
formed the single spectrum approach and is thus the method
of choice in high speed dynamic environments that generate
large amounts of concepts over the progression of the stream.
In such environments FCT would be challenged in terms of
memory capacity and would be forced to flush portions of
its repository sooner that EP, thus losing its ability to exploit
concept recurrences and in turn leading to a loss of accuracy.
However, as shown in the experimentation care needs to be
taken on how spectra are combined: a naive approach of simply
combining similarly performing spectra in terms of accuracy
can be worse than maintaining single spectra. We showed
that the structural similarity scheme outperformed the other
two approaches on a broad set of criteria including accuracy,
robustness to noise and over-fitting, memory consumption and
processing speed.
In terms of future work there are two promising directions.
We believe that is possible to further reduce the computational
effort involved in deriving the spectrum by only keeping the
lowest order coefficient at each leaf node of the Decision
Tree together with a residual coefficient that captures the
contribution of other coefficients at that node. Secondly, at
each concept drift point we can parallelize computation of the
spectrum in one thread while processing incoming instances
in another thread in a parallel environment such as a Spark
framework. REFERENCES
[1] C. Alippi, G. Boracchi and M. Roveri. Just-In-Time Classifiers for Re-
current Concepts. IEEE Transactions on Neural Networks and Learning
Systems, vol. 24(4), pages 620–634, 2013.
[2] Bifet, A. & Gavald`
a, R. Learning from Time-Changing Data with
Adaptive Windowing. In Proceedings of the 7th SIAM ICDM, pages
443–448. SIAM, 2007.
[3] Bifet, A. Holmes, G. Kirkby, R. & Pfahringer, B. MOA: Massive Online
Analysis. The Journal of Machine Learning Research, vol(11), pages
1601–1604, 2010.
[4] Domingos, P. & Hulten, G. Mining High-speed Data Streams. In
Proceedings of the ACM SIGKDD’00, pages 71–80, New York, NY,
USA, 2000. ACM.
[5] Gama, J. & Kosina, P. Learning about the Learning Process. In Advances
in Intelligent Data Analysis X, vol(7014) of Lecture Notes in Computer
Science, pages 162–172. Springer Berlin Heidelberg, 2011.
[6] Gashler, M., Giraud-Carrier C., & Martinez, T. Decision Tree Ensemble:
Small Heterogeneous Is Better Than Large Homogeneous. 7th Interna-
tional Conference on Machine Learning and Applications, pages 900–
905,IEEE Computer Society, 2008.
[7] Gomes, J. Menasalvas, E. & Sousa, P. Tracking Recurrent Concepts
Using Context. In Rough Sets and Current Trends in Computing,
vol(6086), pages 168–177. Springer Berlin Heidelberg, 2010.
[8] Hoeglinger, S. Pears, R. & Koh, Y. CBDT: A Concept Based Approach
to Data Stream Mining. In Proceedings of the PAKDD ’09, pages 1006–
1012, Berlin, Heidelberg, 2009. Springer-Verlag.
[9] Katakis, I. Tsoumakas, G. & Vlahavas, I. An Ensemble of Classifiers for
Coping with Recurring Contexts in Data Streams. In Proceedings of the
ECAI’08 , pages 763–764, Amsterdam, Netherlands, The Netherlands,
2008. The IOS Press.
[10] Kargupta, H. Park, B. & Dutta, H. Orthogonal Decision Trees. IEEE
Transactions on Knowledge and Data Engineering, vol(18), no(8), pages
1028–1042, 2006.
[11] Lazarescu, M. A Multi-Resolution Learning Approach to Tracking
Concept Drift and Recurrent Concepts. In 5th international workshop
on Pattern Recognition in Information Systems, 2005.
[12] Park, B. Knowledge Discovery from Heterogeneous Data Streams Using
Fourier Spectrum of Decision Trees. PhD thesis, Pullman, WA, USA,
2001.
[13] Pears, R. Sripirakas, S. & Koh, Y. Detecting concept change in dynamic
data streams. Machine Learning, 97:3, pp 259–293, 2014.
[14] Peipei Li, Xindong Wu, and Xuegang Hu, ”Mining recurring concept
drifts with limited labeled streaming data,” ACM Trans. Intell. Syst.
Technol.,vol. 3, no. 2, pp. 29:1-29:32, Feb. 2012
[15] Ramamurthy, S. & Bhatnagar, R. Tracking recurrent concept drift
in streaming data using ensemble classifiers. In 6th International
Conference on Machine Learning Applications, pages 404–409, Dec
2007.
[16] Sripirakas, S. & Pears, R. Mining Recurrent Concepts in Data Streams
Using the Discrete Fourier Transform. In DaWaK’14, vol(8646) of
Lecture Notes in Computer Science, pp 439–451. Springer International
Publishing, 2014.