Incremental learning from unbalanced data with concept class, concept drift and missing features

International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.4, No.6, November 2014
INCREMENTAL LEARNING FROM
UNBALANCED DATA WITH CONCEPT CLASS,
CONCEPT DRIFT AND MISSING FEATURES: A
REVIEW
Pallavi Kulkarni and Roshani Ade
Dr. D. Y. Patil School of Engineering and Technology,
Savitribai Phule Pune University, India
ABSTRACT
Recently, stream data mining applications has drawn vital attention from several research communities.
Stream data is continuous form of data which is distinguished by its online nature. Traditionally, machine
learning area has been developing learning algorithms that have certain assumptions on underlying
distribution of data such as data should have predetermined distribution. Such constraints on the problem
domain lead the way for development of smart learning algorithms performance is theoretically verifiable.
Real-word situations are different than this restricted model. Applications usually suffers from problems
such as unbalanced data distribution. Additionally, data picked from non-stationary environments are also
usual in real world applications, resulting in the “concept drift” which is related with data stream
examples. These issues have been separately addressed by the researchers, also, it is observed that joint
problem of class imbalance and concept drift has got relatively little research. If the final objective of
clever machine learning techniques is to be able to address a broad spectrum of real world applications,
then the necessity for a universal framework for learning from and tailoring (adapting) to, environment
where drift in concepts may occur and unbalanced data distribution is present can be hardly exaggerated.
In this paper, we first present an overview of issues that are observed in stream data mining scenarios,
followed by a complete review of recent research in dealing with each of the issue.
KEYWORDS
Incremental learning, class imbalance, concept class, concept drift, missing features
1. INTRODUCTION
There are various data mining techniques that can be used for analysing problems in real world
applications and discovering their solution in a scientific manner. Supervised learning techniques
use instances, which have already been pre-classified in some manner. That means each instance
has a label, which recognizes the class to which it belongs. Classification is a supervised data
mining technique, predicts about data values, using already known results found from data
gathered in advance. Classification maps, data into groups of classes established in advance. In
short, it tells us which data instance, should belong to which category of data, thereby simplifying
the processing of large amounts of data, such as stream data or data from decision support
systems. Data evolves over time, as time goes on and rapid increase in volume of data is observed
[75; 76].
Conventionally, the supervised learning method attempts to group the data from a static dataset,
the instances of which are related to the basic distribution described by a generating function.
DOI : 10.5121/ijdkp.2014.4602 15

Hence it is presumed that the dataset contains all required information for learning the pertinent
concepts. However, this technique has proven impractical for many real world applications, e.g.,
spam identification, climate monitoring, credit card fraud detection, etc. Data is usually received
over time in streams of instances and is not at hand from the beginning. Such data conventionally
reaches there in two ways, first is incremental arrival or batch arrival. So the question is to use all
the information up to a certain time step t to estimate up to the minute instances coming at time
step t+1. Learning in such cases is referred to as incremental learning [35].
If learning from enormous data which is progressing by sequential steps is desired, then
incremental learning or online algorithms are best suitable and preferred. Incremental learning
algorithm is always fed with input data as the data arrives in sequence, the algorithm computes
hypothesis by taking into account order of data arrival, each one gives details of all the data seen
so far. Note that the algorithm should be depend on prior hypotheses and the ongoing training
data [66].
16
Incremental learning algorithm should possess all the properties mentioned below:
1. The algorithm should gain knowledge which is additional in recent data.
2. To train already present classifier, the algorithm need not access to the initial data
3. It should keep up the knowledge which it has previously learned
(No catastrophic forgetting should occur)
4. New data may bring in concept class. The algorithm should be able to recognize and
accommodate such a new class which can be introduced over the period of time. [1]
This definition points out that, learning from stream data needs a classifier that can be
incrementally modified to get full benefit from newly arrived data, while simultaneously
preserving performance of the classifier on matured data underlining the fact to of “stability-plasticity”
dilemma which describes how a learning system can be designed to stay stable and
unchanged to immaterial affairs , while plastic (i.e. be able to change when necessary in order to
deal with different situations) to recent and crucial data (e.g. concept change) .
Hence, the stability-plasticity dilemma poses a cline on which incremental learning classifiers can
prevail. Batch learning algorithms (i.e. system of algorithms trained on static data) are on one end
of the cline which is representation of stability. These algorithms overlook al new data, instead is
aimed fully at formerly learned concepts.
Another end of the cline contain online learning algorithms where the model is altered upon
observing the new instance and the instance is discarded immediately.
Even if batch learning systems prevail on one of the end of cline of stability-plasticity dilemma,
they are fundamentally not incremental in nature. Because batch learners do not have the capacity
of describing any new instances once they have been learned and thus fail to fulfill property 1 of
incremental learning. This restraint is alleviated by building ensembles of batch learners, where
fresh batch learners can be learned on the fresh data, and then united through a voting system [28;
74].
Ensemble technique (multiple classifier system) is widespread in the area of machine learning,
specifically in the incremental environment. An ensemble technique is extracted by merging
diverse classifiers. There are various differentiating parameters who help to achieve diversity that
in turn entitles each classifier to produce several decision boundaries. Appropriate diversity
allows to gain different errors to be made by individual classifier and finally strategic integration
of them can cut off the total error in the entire system. The Ensemble can be build up in several
ways like:

17
1) Bagging
2) Boosting – Adaboost is pretty popular algorithm
3) Stacked generalization
4) Mixture of experts
The diversity need can be fulfilled by applying various approaches such as:
1) Training each classifier using several data chunks
2) Training each classifier using several parameters of a given classifier architecture
3) Training each classifier using several classifier models
4) Random Subspace method (training each classifier using several subset of characteristics)
Detail literature survey on the classifier combination approach can be found in Kuncheva’s book
[83].
Ensemble approach can be used to address every challenge associated with incremental learning
as a basic design model. Several solutions for those issues are found that are inherently combined
with ensemble systems [78-83].
2. ISSUES OF DATA STREAM LEARNING – CONCEPT CLASS
One of the most important work in incremental learning includes Learn++ group of algorithms.
The initial algorithm from Learn++ family is Learn++ which is inspired by Adaboost[66].
Similarities between them are:
1) Sequential generation of classifier ensemble
2) Training data contains bootstrapped samples which are taken from distribution
3) Weight Distribution is altered iteratively
The main difference between them is in the working of distribution update rule, where this rule is
better in terms of performance in case of Learn++ as the distribution moves step by step towards
the new instances that have not been appropriately learned by the present ensemble in the system.
Classifiers are united through weighted majority voting technique, where voting weights are
computed by relative performance (in terms of minimum error) of each classifier on instances of
training data. If new data arrives, Learn++ creates additional ensemble to learn novel information
from it. Learning concept classes need large number of classifiers (referred to as problem of
classifier proliferation) for each novel class to be learned in Learn++. So it is possible to learn
new concept classes in Learn++ but it does so with added cost [1].
The interesting quality in Learn++. NC is a novel classifier combination technique that lets
distinctive classifiers to communicate with each other to decide their voting weights for each test
example. Learn++. NC applies different set of rules for updating initialization of the algorithm
when recent data comes. Learn++.NC enquire each specific classifiers to cross check their results
with respect to classes on which training was given to them. By observing the predictions of
another classifiers, each classifier determines whether its result is in line with the classes others
are deciding, and the training classes. If this is not the case, then the classifier cut down its vote,
or possibly desist from voting all together with all other classifiers in the ensemble system. Hence
problem of out-voting is frankly discussed and addressed in Learn++.NC (New Class), through a
special mechanism referred to as dynamic consult and vote [28]. Learn++.NC cannot address the
infrequent but difficult to control process of adding new classes and at the same time discarding
former ones on a next dataset Learn++.UDNC is an improvement over Learn++.UD [64] and
scrounges confidence measure mechanisms from Learn++.NC. It is discussed in next sections of
the paper [52].

Gregory et al. proposed a smart hybrid algorithm which is part of Learn++ family and referred to
as Learn++.NCS (New Classes with Sampling).It uses a SMOTE as wrapper to make process of
learning easier of a minority classes even when former/later classes are discarded/added
sporadically in the interval of a testing. The wrapper can be applied to other learning algorithms
than Learn++.NC. In other words, Learn++.NCS uses Learn++.NC for the incremental learning
of concept classes along with a SMOTE-based sampling wrapper to handle class imbalance [65].
He et al. put forward a universal adaptive incremental learning framework called ADAIN which
is having capacity to learn from stream (continuous) data, amassing knowledge over time, and
employing such experience to enhance later learning and prediction (classification) performance.
ADAIN framework has capability to handle and detect single or multiple new concept classes. It
considers input data stream as raw data on which processing has to be done by the distribution
function and finally hypothesis will be generated. This framework also includes a mapping
function whose aim is to stipulate a quantitative judgment of the learning proficiency of the recent
piece of data (note that the size of this piece of data will have crucial influence on the eventual
learning results) based on formerly trained classifier. Additionally this mapping function also
referred to as non-linear regression model, puts up a link from previous wisdom to the recently
accepted data, and adjusts (i.e. quality of being adaptable) such experience to data sets received
later. In this paper, two major design strategies are simulated of the given framework known as
ADAIN.MLP and ADAIN.SVR. They are distinguished by the basic implementation of the
model they are using inside the mapping function. By observing performance of ADAIN
comparatively with other existing approaches, Learn++ algorithm is the most assertive one.
Learn++ is set on the “Ensemble-of-Ensemble” technique. Due to use of this “Ensemble-of-
Ensemble” policy, Learn++ tends to acquire much more computational resources. Additionally,
when the data set is of huge size, the required time of Learn++ increases exponentially, which
restricts its ability to scale and ability to tackle large-scale stream data [86-87].
Therefore, we can say that ADAIN achieves the goal of incremental learning as it successfully
unites formerly learned experience into recently accepted data to enhance learning from current
raw data, and in the meantime it accumulates such knowledge over time to encourage decision-making
in future. However it does not address issues of concept shifting and class imbalance [69-
73].
Several classification simulations can be used as base classifiers in multiple classifier system.
Each classifier gets a chunk of data distribution in the form of some signals or we can call it as a
pattern.
This pattern describes how the system is going to operate and once applied to the ensemble of
classifiers can generate output in the form of fault. By considering his idea, Roozbeh et al.[68]
proposed a new method where the MultiLayer Perceptron (MLP) from Neural Networks[67] are
used as the base classifier in the ensemble system which can in turn be used for training of
ensemble and testing of unseen instances further. Three layer MLP are employed because they are
able to differentiate between complex and nonlinear decision boundaries and also are adaptable of
behaving like weak learners by sufficient parameter updation mechanism. Here, the detection of
unknown classes in future data is computed by thresholding the normalized weighted average of
outcomes (NWAO) of the base classifiers in the ensemble system. Noticing of new fault classes is
crucial in applications such as nuclear industry systems. For that purpose, threshold analysis with
NWAO is proposed which sets some limits on the values and thereby simplifies the task of
detecting fault classes in nuclear power system application. Class assignment is performed with
the help of DW-CAV routine of Learn++.NC algorithm [28] in this work.
18

19
3. ISSUES OF DATA STREAM LEARNING – CONCEPT DRIFT
Data is always generated by some function. Conventional system of data mining algorithms
assumes that each dataset is produced from a specific, static, hidden function. It means that
uniform function is used for training and testing of the data. This assumption may fail in case of
stream data, i.e. the function which produces instances at time step t may not be the identical
function as the one that creates instances at time step t+1 (next time instance).This possible
variation in the underlying function is referred to as concept drift. In other words, concept drift
might be viewed in a more abstract sense as a hindrance caused by inadequate, unknown or
unobservable characteristics in a dataset, an event called as hidden context [24]. Here, the
underlying phenomenon which lends a real and static picture over time for each class is
regrettably concealed from the learner’s vision. New data is generated by some hidden function
and learner is not aware of this fact and therefore the concept drift is unpredictable. If the
generating function for the drifting concepts was already recognized, one could simply learn a
suitable classifier for each pertinent concept, and use the appropriate classifier for all recent data
(which is called as the multitask learning problem). In the unavailability of such knowledge, then,
we should configure a unified classifier that is able to acquire such swings in concepts over time.
Diagram below shows the relationship among concept drift, model adaptation, knowledge
(wisdom) sign-over and time series analysis. It is quite clear from the figure that all this features
are dependent on time.
Figure 1. Relation of concept drift with its features
3.1 Formal definition of Concept drift
Concept drift occurs in an environment which is commonly referred to as Non stationary
environment. In such cases we assume that at time step t, the algorithm A is presented with a set
of labeled instances{X0,….,Xt}, where Xi is a v-dimensional vector and each instance has
corresponding class label yj..If an unlabeled instance comes at time t+1 as Xt+1, then the algorithm
is expected to provide class label for Xt+1. This is a predicted label of given instance. Once this is
done, the real label Yt+1 and a new testing instance Xt+2 comes so that we can go for it’s testing.
Moreover, we call the hidden function fh producing the instance at time instant t as ft. Concept
drift is said to come up when the underlying generating function (fh) alters over time. There are
many styles in which this alteration can take place. Bayes’ theorem can calculate the probability
that instance Xt+1 belongs to class ci as:
This well- known theorem gives relationship between prior and posterior probabilities of
instances and their classes.

20
Concept drift can take place with respect to any of the three main variables in Bayes’ theorem
1) p(ci) can alter(prior class distribution)
2) p(Xt+1|c) can alter (class distribution)
3) p(c|Xt+1) can alter (posterior class distribution)
There are also several forms of concept drift which can be categorized into two types as, real
concept drift and virtual concept drift.[22,23] If the distribution of instances is altering
corresponding to alteration in the class prior probabilities or change in class distribution is taking
place then virtual concept drift is said to occur. In this case underlying concept means posterior
distribution of instances remains same as before. This may create problems for the learner, as
such updates in the probability distribution may alter the error of the learned system, even if the
concept did not alter. Some parts of target concepts may have gone unobserved by the learner,
because of alteration of distribution such instances may become more common Sine the learner
never seen such data, it could not learn the concept and hence should be retrained. This kind of
virtual drift is particularly applicable when the data stream displays unbalanced classes.
Also, real concept drift can be interpreted as a modification in the class boundary (i.e. change in
the posterior distribution). This alteration indicates more fundamental change in the generating
function.
Note that real concept drift need alteration in the model and virtual concept drift also require the
same. This leads to the fact that final result is same. Moreno-Torres et. al. [21] stipulate clearer
survey of various forms of “dataset drift”, as well as their Sources.
Quinonero-Candela [25], Minku [26], and Kuncheva [27], [28] stipulated complete summaries for
featuring several types of concept drift with respect to its speed, randomness, and cyclical nature.
Drift speed is the displacement rate in prior probability of an instance from one time step to the
later one. Substantial displacement within a step represents fast (rapid) drift and usually results in
high classifier error. Gradual (slow) drift, however, sounds in compact displacements which in
turn leads to lower classification error, and as a result, is harder to notice.
Drift randomness is a significant descriptor in being able to show good judgment about quality
between non-stationary and noisy data, and can be illustrated as the change in a distribution over
a short spell of time. Randomness can be viewed in terms of its frequency and magnitude: high
variance between two spells of time denotes a highly unstable environment which, as this level
raises, reaches a condition where the environment cannot gain new knowledge. The cyclical
nature of drift is an event that can be seen in different real-world applications such as weather
forecasting, climate monitoring, nuclear power claim modeling. In such situations, class
definitions alter in such a way that a last settings of the environment may happen again or number
of times (i.e. recur) after some spell of time. Recurrence which can be seen in such environments
is one of a two kind as periodic or random. Concept drift learning systems do not address the
problem of concept class i.e. addition or removal of new classes; as such phenomenon are
featured more by Concept change.
Therefore speaking in the more general terminology such learning in nonstationary environments
or nonstationary learning (NSL) is to refer to any drift or change without considering its kind and
nature. [2]
There are some general guidelines to shape up a system for learning in non-stationary
environments.

1) Necessity of truly incremental or one pass learning where access to previous data is strictly not
allowed for future training. It means that any instance is processed only once with training and at
that time only the knowledge must be retrieved and summarized so that it can be used in model
building process [1]
2) It is known fact that the most recent dataset is a portrayal of the present environment. So the
knowledge must be grouped based on its relevance to the present environment, and be
dynamically brought up to date as latest data arrives.
3) The learner should have a system to resign itself when former and recently learned knowledge
dispute with each other. Moreover, there should be a system for keeping track of both the
incoming data and the learner’s performance on recent and existing data for the intention of
complexity reduction, problematizing, and fading.
4) The learner must have a system to neglect or omit information that is no longer relevant, but
selectively with the added capability to look back such information if the drift or change chase a
cyclical nature.
5) Knowledge should be incrementally stored after certain chunk of time interval so that it can
start working to generate the wise speculation for an unknown (or unlabeled) data instance
through any instant of time in the learning routine.[29]
21
3.2 Existing approaches for handling concept drift
Concept drift algorithms can be categorized in different ways, like:
1) Online vs. batch approaches
2) Single classifier vs. ensemble-based approaches
3) Incremental vs. non-incremental approaches
4) Active vs. passive approaches
Various concept drift learning algorithms exploit some kind of a sliding window over the
incoming data, where the group of instances that fall within the window are assumed to be
stationary, and a novel classifier is produced for each such group of data. STAGGER [3] and
FLORA [4] are the very first systems who used this passive batch-based instance selection
method. Some versions of FLORA algorithms consist of an active drift detection system, using an
adaptive window narrowing or widening depending on whether the drift is fast or gradual [4].
FLORA calls the classifiers as relevant, irrelevant or potentially relevant by assessing them on the
most current data. Each classifier sustains a counter based on the number of accurately classified
instances, and classifiers are cut off based on their relevance in the present data window. But such
an approach can result in catastrophic forgetting [5]. Another bunch of active concept drift
methods make use of control charts, such as CUSUM (cumulative sum). Alippi and Roveri’s just-in-
time (JIT) classifiers [6-8], and their more latest intersection of confidence intervals (ICI) rule
[9] are cases of such methods. Information theoretic measures, such as entropy, mutual
information, or Hoeffding bounds of individual features have also been used for noting drift and
changing a classifier like a decision tree [10-12]. Several methods among them also consists of a
FLORA-like windowing system, including Hulten et al.’s concept adapting very fast decision tree
(CVFDT) [13] or Cohen et al.’s incremental online-information network (IOLIN) algorithms
[14;15].
Most of the methods of drift detection are generally fairly able to gain success in detecting curt
changes, but may clash with slow drift. The Early Drift Detection Method (EDDM) is particularly

designed for slow drift as an online active method[16]. It keeps an eye on the distance between
the errors of a classifier and compares their mean to a threshold. EDDM not only indicates the
drift, but also it has the capability to raise a warning if required.
22
3.3 Ensemble techniques as a solution for concept drift
Ensemble systems are also called as multiple classifier systems (MCS), have been designed and
successfully carried out for learning in non-stationary environments [17]. With each and every
new dataset arrived, new classifiers are added and many times older one are removed to create an
ensemble model to keep track of the environment.
These methods generally go for a passive drift detection along with a fixed ensemble size
(number of classifiers), where the most former member (e.g. Street’s Streaming Ensemble
Algorithm (SEA) [18], and Bifet’s adaptive Hoeffding tree bagging [19]) or the ensemble
member which gives minimum performance (e.g.Tsymbal’s Dynamic Integration [20], Kolter and
Maloof’s, Dynamic Weighted Majority (DWM) [30]) is substituted with a recent one.
Although there are disagreements on which scheme to use for concept drift applications, for
merging classifiers voting mechanism is preferred quite often.
Tsymbal merges a proximity measure with classifier performance to compute the voting weights,
with classifiers whose training and test data are in the identical region of feature space being
offered more weights [20]. Gao, recommends simple majority voting [33], highlighting that
weights based on classifier error on drifting data is not giving enough information for later
datasets. Other versions of ensemble methods consists of [34-36].
Hybrid methods that combine active detection, sliding windows and classifier ensembles have
also been stated, such as random forests with entropy [37] and Bifet’s novel integration of a
Kalman filter merged with an adaptive sliding window algorithm, namely ADWIN [38].
Bifet et. al. has launched a software suite similar to WEKA, called Massive Online Analysis
(MOA) at [39], which contains implementations of ADWIN and a various other tools for stream
data mining with concept drift. Another application where active ensemble method id
implemented is Hoens et. al.’s latest work that merges random subspace approach with Hellinger
distance to observed drift in traits of the data [44].
Ensemble based concept drift algorithms also refer to some of the algorithms from Learn++
family such as the incremental Learn++.NSE algorithm [29;41-44], which produces a novel
classifier with each batch of data that arrive. The classifiers are merged via dynamically weighted
majority voting. Weights are developed from the time-adjusted error of each and every classifier
in the ensemble, sigmoidally averaged over all environments. Learn++.NSE is able to keep track
of several environments, such as gradual, fast, abrupt, and cyclical drift. It is able to recognize
most and least relevant classifiers, give them high and low voting weights, and notice when a
classifier turn relevant again, whether the environment and its settings peruse a cyclic nature. The
algorithm is also able to take in class which are added or removed. Only thing it cannot do is it is
not designed to handle class imbalance.
4. ISSUES OF DATA STREAM LEARNING – CONCEPT IMBALANCE
Class imbalance happens when a dataset does not contain (approximately) identical number of
instances form each class, which can be drastic in many situations [45].Imbalanced data learning
is a major issue and its detailed survey with assessment matrix and state-of-art techniques to

address the problem is available in [46]. A simple solution to this problem is under/over sampling
of majority/minority class data. But this approach has pitfalls: under-sampling discards instances
from the majority class, without considering usefulness of those instances. On the other side,
oversampling generates duplicates of the minority examples, which might turn classifier to overfit
the minority class examples. There are smarter mechanisms like condensed nearest neighbor
(CNN) rule which discards instances from the majority class that are remote from the decision
boundary [47]. In one more approach, Tomek joins that under-samples the majority class by
clarifying a distance between examples from various classes [48].
An intelligent approach, is used by Chawla’s SMOTE (Synthetic Minority Oversampling
TEchnique) algorithm [49], which inhabits the minority class feature space by tactically putting
artificial instances on the line segment connecting two minority examples. SMOTE has been
proved to enhance the classification accuracy on the minority class over other standard
mechanisms. Also, SMOTEBoost [50] merges SMOTE and AdaBoost.M2 to further improve F-measure
and recall. Bagging ensemble variation (BEV) was launched [51], which trains
classifiers by applying bagging method on entire minority class data and subsets are taken for
training from majority class data. Learning from unbalanced data is an issue which was addressed
in an algorithm called Learn++.UD [64] but it is not having capability to learn new concept
classes. Learn++.UDNC is joint solution to the problem of unbalanced data as well as learning
from new concept class. In other words it is capable to incrementally learn latest concept classes
from unbalanced datasets. Alternatively, Learn++.UDNC collectively implements new qualities
of several algorithms within the Learn++ family, like a class specific weighting approach,
normalized preliminary confidence measures and introduces new transfer function that is able to
bring down the confidence bias of a sub-ensemble which is trained on a majority class[52].
DataBoost-IM algorithm discovers difficult to classify minority class examples, based on which
generates new artificial data [53]. Classifier precise methods, like rebalancing planned to deal
with support vector machines (SVMs), have also been launched recently [54; 55].
23
5. ISSUES OF DATA STREAM LEARNING –JOINT PROBLEM OF CLASS
IMBALANCE AND CONCEPT DRIFT
A massive amount of applications in non-stationary environments with concept drifting data
sources are affected from class imbalance (e.g. Climate monitoring, spam e mail identification,
credit card fraud detection, network intrusion detection system).Ensemble approaches are the one
which are mostly used when handling concept drift, in addition they have been proved useful for
combating class imbalance [36].
A learning algorithm for non-stationary environment from imbalanced data has been stated by
Gao et al. [36; 57]. The algorithm called as uncorrelated bagging (UCB), is based on a bagging
framework that trains classifiers on a subset of the majority class instances (which is selected by
some user defined threshold parameter) and the combination of all minority class examples
observed so far (present + former positive instances). There are several drawbacks of this
approach due to its implicit assumption. It is not a true one pass (incremental) algorithm either.
Chen and He’s Selectively Recursive Approach (SERA) learning algorithm goes for a similarity
measure to choose former minority examples that are most close to those in the most recent
dataset [58]. SERA is less liable to challenges of the minority class drifting compared to UCB
[60], and may be implemented in either of two methods: i) create only one classifier on each
dataset; or ii) Apply biased bagging, BBagging, which itself by its own raises number of sampling
weights of the minority data. SERA omits instances from the present training set by what it
considers as not useful by using a Mahalanobis distance measure. SERA is updated to apply an
ensemble technique in [59; 60]. But, this technique is also not strictly single pass, as it accesses
previous data.

Therefore both approaches written above perform best when the minority data concept is
stationary and/or the former (minority class) data can be stored for later use. Xioufis et. al. came
up with a window-based approach that employs a k-NN for multi-label classification for data that
includes issues like concept drift and class imbalance [61].
Ditzler and Polikar proposed two ensemble based techniques that can learn in a broad area of
concept drift in nonstationary environment that addresses heavy class imbalance problem, also
avoids the main limitation of previous methoss, namely, amassing minority data and employing
some portions of former data [62]. The approaches known as Learn++.NIE and Learn++.CDS,
are truly incremental (one pass) methods that do not need access to former data, and they do not
build up minority data to balance the class balance. Smartly, Learn++.CDS applies SMOTE to
rebalance the classes and its instances, whereas Learn++.NIE employs sub-ensembles with
bagging, along with alternate error measures to learn from imbalanced data. Both algorithms are
capable to gain recent knowledge and save old information about the environment and its settings,
which is specifically useful for recurring concepts (e.g. spiral dataset). Particularly, Learn++.CDS
is a union of the two algorithms, Learn++.NSE [7] and SMOTE [46], existing methods for
concept drift and imbalanced data issues, respectively. The robust approach used SMOTE to
rebalance the classes employing artificial (synthetic) minority class data, also used Learn++.NSE
on rebalanced data.
Learn++.NIE, is more intelligent approach which employs a different technique that is mainly
based on two pillars: 1) Employ bagging based sub-ensembles to cut off class imbalance (without
producing synthetic data, and without collecting minority data); and 2) Apply different measures
that gives stress on both class-specific performances to weigh classifiers.
Learn++.NIE algorithm performs well, if both minority and majority class definitions are
changing and a strong balancing is required on both minority and majority classes to achieve
better performance . It has the unique characteristic to hold performance for class specific recall
[63].
24
6. ISSUES OF DATA STREAM LEARNING –MISSING FEATURES
The unity and completeness of data are necessary for any classification system. A trained
classifier requires special training to address this challenge and cannot operate examples with lost
characteristics. Missing data in real world applications is a common scenario: malfunctioning
sensors, wrong pixel information, blank answers to questions asked in surveys, failed equipment,
medical tests that cannot be monitored under some specific situations, etc. are all common
applications in real-life that can result in lost characteristics. Values of features that are beyond
certain dynamic minimum and maximum limit of the data caused by utmost noise, signal
saturation, data corruption, etc. can also be considered as missing features. The simplest way to
deal with such lost data is to neglect those instances that have missing attributes. When large
portions of the data is facing problem of missing features or lost attributes filtering or deletion
(list wise) are suboptimal (and impractical due to conditions applied on them to work well)
approaches and generally known as filtering [85]. A more realistic approach is imputation where
missing values are filled with a meaningful calculation [86-89] like mean or finding k-nearest
neighbor value etc. to get correct estimates of missing values training data should be dense with
enough number of features.
These techniques tends to produce biased results. Polynomial regression is one more method to
handle missing data but regression technique is suitable to some specific kind of applications.
Approaches with good performance guarantees have also been discovered. Many of these
approaches are based on model based estimation, like Bayesian estimation [90-92], these method

calculate posterior and prior probabilities by uniting the missing feature space. Such methods also
need enough dense data distribution and data distribution with parameters should be known in
advance. Such prior knowledge in difficult to obtain in most of the cases. Alternatively
Expectation Maximization (EM) [93-95; 91] algorithm is theoretically proven iterative method
which is even simple to construct. EM suffers from two limitations: 1) Slow convergence if large
part of data has missing values; and 2) Maximization step may be pretty hard if a closed form of
the data distribution is not known, or if different examples have lost different features. So in such
cases, EM does not work well in real world [2]. EM also needs information in advance about the
underlying distribution that is not available usually.
If the distribution calculation is wrong then it may lead to inconsistent outputs, whereas
unavailability of enough dense data may cause accuracy loss. To overcome such difficulties,
rigorous approaches have been suggested such as using Gaussian mixture models [96, 97]; or
Expected conditional Maximization [86]. There are several other approaches such as neural
network based methods [98, 99], neuro –fuzzy algorithms [100]. Algorithms based on the
common fuzzy min-max neural network or ARTMAP and fuzzy c-means clustering [101] are
some of the examples of the method. Ensemble based approaches have also been launched such
as Melville et al. [102] proved that the algorithm DECORATE that produces artificial data (with
lack of missing values) from already present data (with the presence of missing values) is pretty
robust to lost features. Juszczak and Duin [103] suggested uniting an ensemble of one class
classifiers, where each classifier trained on one feature. This method is able to handle any type of
combination of missing data, with less number of classifiers. This technique can be much
powerful as long as single characteristic is able to calculate the underlying decision boundaries.
But that is not often reasonable. Polikar et al. proposed a new way that belongs to Learn++ family
known as Learn++.MF [84] produces sufficient frequency of classifiers whereby each one is
trained with randomly chosen feature subset.
An instance with lost values is classified by majority voting of those classifiers that didn’t
employ the lost values in the period of the training. Therefore, this method can be distinguished
by other methods in fundamental aspect. Learn++.MF attempts to extract the most unfair
classification information stipulated by the present (old) data, so that it can take most benefit of
duplication in the feature set. In other words, Learn++.MF do not have any concern with the
estimation prediction of missing data values. Therefore, Learn++.MF prevents many of the
demerits of estimation and imputation based approaches. This is a novel way to accommodate
crucial amount of lost data values and without slow decline in performance as the volume of data
grows. A detailed analysis can be found in [84] with simulation results and conditions in which
the given algorithm performs at its best.
25
7. CONCLUSION AND FUTURE WORK
In this paper, we arranged our survey in terms of the four main challenges of research in handling
stream data. We commented on the major areas of research in mining data streams that exhibit
concept class, concept drift, class imbalance, and missing features. We discussed adaptive
approach to handle concept class, combine problem of class imbalance and concept drift, and
various techniques to solve such problems including ensemble technique. We observed that
Learn++ family of algorithm play in important role in this problem domain along with adaptive
framework known as ADAIN.
In spite of the increasing number of attempts, there is still much work to be done in data streams
that constitute concept drift, class imbalance and concept class. We note that “ADAIN” solution
of incremental learning is quite intelligent but does not handle concept drift and imbalance.
Learn++.NIE algorithm is elegant technique for addressing class imbalance in non-stationary

environment. The work in this area with handling of this issues jointly is sparse and have not been
evaluated thoroughly on large and real world applications. Thus we propose future work leading
towards an intelligent approach that will overcome limitations of the existing research body and
also rigorously assessing these techniques on large scale, real world examples.
REFERENCES
[1] R. Polikar, L. Udpa, S. S. Udpa, and V. Honavar, “Learn++: An incremental learning algorithm for
supervised neural networks,” IEEE Trans. Syst., Man Cybern. Part C: Appl. Rev., vol. 31, no. 4, pp.
497– 508, Nov. 2001.
[2] R. Elwell and R. Polikar, "Incremental Learning of Concept Drift in Nonstationary Environments,"
26
IEEE Transactions on Neural Networks, vol. 22, no. 10, pp. 1517-1531, Oct.2011
[3] J. C. Schlimmer and R. H. Granger, "Incremental learning from noisy data," Machine Learning, vol.
1, no. 3, pp. 317-354, Sept.1986.
[4] 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.
[5] F. H. Hamker, "Life-long learning Cell Structures--continuously learning without catastrophic
interference," Neural Networks, vol. 14, no. 4-5, pp. 551-573, May2001.
[6] C. Alippi and M. Roveri, "Just-in-Time Adaptive Classifiers - Part I: Detecting Nonstationary
Changes," IEEE Trans. Neural Networks, vol. 19, no. 7, pp.1145-1153, 2008.
[7] C. Alippi and M. Roveri, "Just-in-Time Adaptive Classifiers - Part II: Designing the Classifier," IEEE
Transactions on Neural Networks, vol. 19, no. 12,pp. 2053-2064, Dec.2008.
[8] C. Alippi, G. Boracchi, and M. Roveri, "Just in time classifiers: managing the slow drift case,"
International Joint Conference on Neural Networks (IJCNN 2009), pp. 114-120, Atlanta, GA, 2009.
[9] C. Alippi, G. Boracchi, and M. Roveri, "Change Detection Tests Using the ICI Rule," World
Congress on Computational Intelligence (WCCI 2010) - International Joint Conference on Neural
Networks (IJCNN 2010), pp. 1190-1196, Barcelona, Spain, 2010.
[10] P. Vorburger and A. Bernstein, "Entropy-based Concept Shift Detection," International Conference
on Data Mining (ICDM '06.), pp. 1113-1118, 2006.
[11] S. Hoeglinger and R. Pears, "Use of Hoeffding trees in concept based data stream mining,"
International Conference on Information and Automation for Sustainability (CIAFS 2007), pp. 57-62,
2007.
[12] C. J. Tsai, C. I. Lee, and W. P. Yang, "Mining decision rules on data streams in the presence of
concept drifts," Expert Systems with Applications, vol. 36, no. 2, Part 1, pp. 1164-1178, Mar.2009
[13] G. Hulten, L. Spencer, and P. Domingos, "Mining time-changing data streams," Conf. on Knowledge
Discovery in Data, pp. 97-106, San Francisco, 2001
[14] L. Cohen, G. Avrahami, M. Last, and A. Kandel, "Info-fuzzy algorithms for mining dynamic data
streams," Applied Soft Computing, vol. 8, no. 4, pp. 1283-1294, Sept.2008
[15] L. Cohen, G. Avrahami-Bakish, M. Last, A. Kandel, and O. Kipersztok, "Real-time data mining of
non-stationary data streams from sensor networks," Information Fusion, vol. 9, no. 3, pp. 344-353,
2008
[16] M. Baena-Garcia, J. del Campo-Avila, R. Fidalgo, A. Bifet, R. Gavalda, and R. Bueno-Morales,
"Early drift detection method," ECML PKDD Workshop on Knowledge Discovery from Data
Streams, pp. 77-86, 2006.
[17] L. I. Kuncheva, "Classifier ensembles for detecting concept change in streaming data: overview and
perspectives," European Conference on Artificial Intelligence (ECAI), pp. 5-10, Patras, Greece, 2008.
[18] W. N. Street and Y. Kim, "A streaming ensemble algorithm (SEA) for large-scale classification," Int'l
Conf. on Knowledge Discovery & Data Mining, pp. 377-382, 2001.
[19] A. Bifet, G. Holmes, B. Pfahringer, R. Kirkby, and R. Gavalda, "New Ensemble Methods For
Evolving Data Streams," 15th ACM SIGKDD international conference on knowledge discovery and
data mining (KDD 09), pp. 139-148, 2009
[20] A. Tsymbal, M. Pechenizkiy, P. Cunningham, and S. Puuronen, "Dynamic integration of classifiers
for handling concept drift," Information Fusion, vol. 9, no. 1, pp. 56-68, Jan.2008.
[21] Tsymbal, A.: The problem of concept drift: definitions and related work. Tech. Rep. TCD-CS-2004-
15, Departament of Computer Science, Trinity College (2004).
URL https://www.cs.tcd.ie/publications/techreports/reports
[22] Widmer, G., Kubat, M.: Effective learning in dynamic environments by explicit context tracking. In:
ECML, pp. 227–243. Springer (1993)

[23] G. Widmer and M. Kubat, “Learning in the presence of concept drift and hidden contexts,” Mach.
27
Learn., vol. 23, no. 1, pp. 69–101, 1996
[24] J. Quinonero-Candela, M. Sugiyama, A. Schwaighofer, and N. D. Lawrence, Dataset Shift in Machine
Learning. Cambridge, MA: MIT Press, 2009
[25] L. L. Minku, A. P. White, and Y. Xin, “The impact of diversity on online ensemble learning in the
presence of concept drift,” IEEE Trans. Knowl. Data Eng., vol. 22, no. 5, pp. 730–742, May 2010.
[26] L. I. Kuncheva, “Classifier ensembles for detecting concept change in streaming data: Overview and
perspectives,” in Proc. Eur. Conf. Artif. Intell., 2008, pp. 5–10.
[27] L. I. Kuncheva, “Classifier ensembles for changing environments,” in Multiple Classifier Systems,
vol. 3077. New York: Springer-Verlag, 2004, pp. 1–15
[28] M. Muhlbaier, A. Topalis, and R. Polikar, “Learn++.NC: Combining ensemble of classifiers with
dynamically weighted consult-and-vote for efficient incremental learning of new classes,” IEEE
Trans. Neural Netw., vol. 20, no. 1, pp. 152–168, Jan. 2009.
[29] R. Elwell and R. Polikar, "Incremental Learning of Concept Drift in Nonstationary Environments,"
IEEE Transactions on Neural Networks, vol. 22, no. 10, pp. 1517-1531, Oct.2011.
[30] J. Z. Kolter and M. A. Maloof, "Dynamic weighted majority: an ensemble method for drifting
concepts," Journal of Machine Learning Research, vol. 8, pp. 2755-2790, 2007
[31] A. Tsymbal, M. Pechenizkiy, P. Cunningham, and S. Puuronen, "Dynamic integration of classifiers
for handling concept drift," Information Fusion, vol. 9, no. 1, pp. 56-68, Jan.2008.
[32] J. Z. Kolter and M. A. Maloof, "Dynamic weighted majority: an ensemble method for drifting
concepts," Journal of Machine Learning Research, vol. 8, pp. 2755-2790, 2007.
[33] J. Gao, W. Fan, and J. Han, "On appropriate assumptions to mine data streams: analysis and practice,"
International Conference on Data Mining (ICDM 2007), pp. 143-152, 2007.
[34] K. Nishida and K. Yamauchi, "Adaptive Classifiers-Ensemble System for Tracking Concept Drift,"
International Conf.on Machine Learning and Cybernetics, eds. K. Yamauchi, Ed., vol. 6, pp. 3607-
3612, 2007.
[35] H. He and S. Chen, "IMORL: Incremental Multiple-Object Recognition and Localization," IEEE
Transactions on Neural Networks, vol. 19, no. 10, pp. 1727-1738, 2008.
[36] J. Gao, B. Ding, F. Wei, H. Jiawei, and P. S. Yu, "Classifying data streams with skewed class
distributions and concept drifts," IEEE Internet Computing, vol. 12, no. 6, pp. 37-49, 2008.
[37] H. Abdulsalam, D. Skillicorn, and P. Martin, "Classification Using Streaming Random Forests," IEEE
Transactions on Knowledge and Data Engineering, vol. 23, no. 1, pp. 22-36, 2011.
[38] A. Bifet, E. Frank, G. Holmes, and B. Pfahringer, "Accurate ensembles for data streams: Combining
restricted Hoeffding trees using stacking.," 2nd Asian Conference on Machine Learning in Journal of
Machine Learning Research, vol. 13, Tokyo, 2010.
[39] A. Bifet, MOA: Massive Online Analysis, Available at: http://moa.cs.waikato.ac.nz/ . Last
accessed:7/6/2011.
[40] T. R. Hoens, N. V. Chawla, and R. Polikar, "Heuristic Updatable Weighted Random Subspaces for
Non-stationary Environments," Data Mining (ICDM), 2011 IEEE 11th International Conference on,
pp. 241-250, 2011.
[41] M. Muhlbaier and R. Polikar, "An Ensemble Approach for Incremental Learning in Nonstationary
Environments," Multiple Classifier Systems, pp. 490- 500, 2007.
[42] M. Karnick, M. Ahiskali, M. Muhlbaier, and R. Polikar, "Learning concept drift in nonstationary
environments using an ensemble of classifiers based approach," Int'l Joint Conf.on Neural Netw., pp.
3455-3462, 2008.
[43] R. Elwell, R. Polikar, "Incremental Learning in Nonstationary Environments with Controlled
Forgetting," International Joint Conference on Neural Net-works (IJCNN 2009), pp. 771-778, 2009
[44] R. Elwell and R. Polikar, "Incremental learning of variable rate concept drift," International
Workshop on Multiple Classifier Systems (MCS 2009) in Lecture Notes in Computer Science, vol.
5519, pp. 142-151, Reykjavik, Iceland, 2009.
[45] M. Kubat, R. Holte, and S. Matwin, "Machine Learning for the Detection of Oil Spills in Satellite
Radar Images ," Machine Learning, vol. 30, pp. 195- 215, 1998.
[46] H. He and E. A. Garcia, "Learning from Imbalanced Data," IEEE Tran. on Knowledge and Data
Engineering, vol. 21, no. 9, pp. 1263-1284, Sept.2009.
[47] P. E. Hart, "The condensed nearest neighbor rule," IEEE Transactions on Information Theory, vol. 14,
no. 3, pp. 515-516, 1968.
[48] I. Tomek, "Two Modifications of CNN," IEEE Transactions on Systems, Man and Cybernetics, Part
A, vol. 6, no. 11, pp. 769-772, 1976.

[49] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, "SMOTE: Synthetic Minority
28
Oversampling Technique," Journal of Artificial Intelligence, vol. 16, pp. 321-357, 2002
[50] N. V. Chawla, A. Lazarevic, L. O. Hall, and K. W. Bowyer, " SMOTEBoost: Improving Prediction of
the Minority Class in Boosting," 7th European Conference on Principles and Practice of Knowledge
Discoveryin Databases (PKDD), pp. 107-119, Dubrovnik, Croatia, 2003.
[51] C. Li, "Classifying imbalanced data using a bagging ensemble variation (BEV)," ACM Southeast
Regional Conf., pp. 203-208, Winston-Salem, NC, 2007
[52] G. Ditzler, M. Muhlbaier, and R. Polikar, "Incremental Learning of New Classes in Unbalanced
Datasets: Learn++.UDNC," Int. Workshop on Multiple Classifier Systems (MCS 2010) in Lecture
Notes in Computer Science, vol. 5997, pp. 33-42, 2010.
[53] H. Guo and H. Viktor, "Learning form Imbalanced Data sets with Boosting and Data Generation: The
DataBoost-IM Approach," ACM SIGKDD Explorations Newsletter, vol. 6, no. 1, pp. 30-39, 2004.
[54] T. Yuchun, Z. Yan-Qing, N. V. Chawla, and S. Krasser, "SVMs Modeling for Highly Imbalanced
Classification," Systems, Man, and Cybernetics, Part B: Cybernetics, IEEE Transactions on, vol. 39,
no. 1, pp. 281-288, Feb.2009.
[55] R. Batuwita and V. Palade, "FSVM-CIL: Fuzzy Support Vector Machines for Class Imbalance
Learning," Fuzzy Systems, IEEE Transactions on, vol. 18, no. 3, pp. 558-571, June2010
[56] Gao, J., Fan, W., Han, J., Yu, P.: A general framework for mining concept-drifting data streams with
skewed distributions. In: SDM, pp. 3–14. Citeseer (2007)
[57] J. Gao, W. Fan, J. Han, and P. S. Yu, "A general framework for mining concept-drifting data streams
with skewed distributions," SIAM International Conference on Data Mining, vol. 7, 2007.
[58] S. Chen and H. He, "SERA: Selectively recursive approach towards nonstationary imbalanced stream
data mining," International Joint Conference on Neural Networks (IJCNN 2009), pp. 522-529,
Atlanta, GA, 2009.
[59] S. Chen, H. He, L. Kang, and S. Desai, "MuSeRA: Multiple Selectively Recursive Approach towards
imbalanced stream data mining," World Congress on Computer Intelligence (WCCI 2010) -
International Joint Conference on Neural Networks (IJCNN 2010), pp. 1-8, Barcelona, Spain, 2010.
[60] S. Chen and H. He, "Towards incremental learning of nonstationary imbalanced data stream: a
multiple selectively recursive approach," Evolving Systems,vol. 2, no. 1, pp. 35-50, 2011.
[61] E. S. Xioufis, M. Spiliopoulou, G. Tsoumakas, and I. Vlahavas, "Dealing with concept drift and class
imbalance in multi-label stream classification," International Joint Conference on Artificial
Intelligence (IJCAI 2011), pp. 1583-1588, 2011.
[62] G. Ditzler, R. Polikar, and N. Chawla, "An Incremental Learning Algorithm for Non-stationary
Environments and Class Imbalance," 20th InternationalConference on Pattern Recognition (ICPR
2010), pp. 2997-3000, Istanbul, Turkey, 2010
[63] Ditzler, Gregory, and Robi Polikar. "Incremental learning of concept drift from streaming imbalanced
data." Knowledge and Data Engineering, IEEE Transactions on 25.10 (2013): 2283-2301.
[64] Muhlbaier, M., Topalis, A., Polikar, R.: Incremental learning from unbalanced data. In: Proc. of Int.
Joint Conference on Neural Networks (IJCNN 2004), Budapest, Hungary, July 2004, pp. 1057–1062
(2004)
[65] Ditzler, Gregory, Gail Rosen, and Robi Polikar. "Incremental learning of new classes from
unbalanced data." Neural Networks (IJCNN), the 2013 International Joint Conference on. IEEE,
2013.
[66] Y. Freund and R. Schapire, “A decision theoretic generalization of on-line learning and an application
to boosting,” Comput. Syst. Sci.,vol. 57, no. 1, pp. 119–139, 1997
[67] C. Bishop, Neural Networks for Pattern Recognition. Oxford. Oxford, U.K.: Oxford Univ., 1995
[68] Razavi-Far, Roozbeh, Piero Baraldi, and Enrico Zio. "Dynamic weighting ensembles for incremental
learning and diagnosing new concept class faults in nuclear power systems." Nuclear Science, IEEE
Transactions on 59.5 (2012): 2520-2530.
[69] He, Haibo, et al. "Incremental learning from stream data." Neural Networks, IEEE Transactions on
22.12 (2011): 1901-1914.
[70] J. Fu, H. He, and X. Zhou, “Adaptive learning and control for MIMO system based on adaptive
dynamic programming,” IEEE Trans. Neural Netw., vol. 22, no. 7, pp. 1133–1148, Jul. 2011.
[71] Ade, Ms RR, et al. "METHODS FOR INCREMENTAL LEARNING: A SURVEY."International
Journal of Data Mining & Knowledge Management Process 3.4 (2013).
[72] H. He, Self-Adaptive Systems for Machine Intelligence. New York: Wiley, 2011

[73] G. G. Yen and P. Meesad, “An effective neuro-fuzzy paradigm for machinery condition health
monitoring,” IEEE Trans. Syst., Man, Cybern., Part B: Cybern., vol. 31, no. 4, pp. 523–536, Aug.
2001
[74] M. Pardowitz, S. Knoop, R. Dillmann, and R. D. Zollner, “Incremental learning of tasks from user
demonstrations, past experiences, and vocal comments,” IEEE Trans. Syst., Man, Cybern., Part B:
Cybern., vol. 37, no. 2, pp. 322–332, Apr. 2007
[75] C. Ji and S. Ma, “Performance and efficiency: Recent advances in supervised learning,” Proc. IEEE,
29
vol. 87, pp. 1519–1535, Sept. 1999
[76] M.D. Muhlbaier, R. Polikar, An ensemble approach for incremental learning in nonstationary
environments, in: 7th International Workshop on Multiple Classifier Systems, Prague, Lecture Notes
in Computer Science, vol. 4472, 2007,pp. 490–500
[77] Hoens, T. Ryan, Robi Polikar, and Nitesh V. Chawla. "Learning from streaming data with concept
drift and imbalance: an overview." Progress in Artificial Intelligence 1.1 (2012): 89-101.
[78] Polikar, Robi. "Ensemble based systems in decision making." Circuits and Systems Magazine, IEEE
6.3 (2006): 21-45.
[79] D. H.Wolpert, “Stacked generalization,” Neural Netw., vol. 5, no. 2, pp. 241–259, 1992
[80] A. Jacobs, M. I. Jordan, S. J. Nowlan, and G. E. Hinton, “Adaptive mixtures of local experts,” Neural
Comput., vol. 3, pp. 79–87, 1991
[81] M. I. Jordan and R. A. Jacobs, “Hierarchical mixtures of experts and the EM algorithm,” Neural
Comput., vol. 6, no. 2, pp. 181–214, 1994
[82] J. Grim, J. Kittler, P. Pudil, and P. Somol, “Information analysis of multiple classifier fusion,” 2nd
Int. Workshop on Multiple Classifier Systems, in Lecture Notes in Computer Science, J. Kittler and F.
Roli, Eds., vol. 2096, pp. 168–177, 2001
[83] L.I. Kuncheva, Combining Pattern Classifiers, Methods and Algorithms. New York, NY: Wiley
Interscience, 2005
[84] Polikar, Robi, et al. "Learn++. MF: A random subspace approach for the missing feature problem."
Pattern Recognition 43.11 (2010): 3817-3832.
[85] Pallavi Kulkarni and Roshani Ade. Article: Prediction of Student's Performance based on Incremental
Learning. International Journal of Computer Applications 99(14):10-16, August 2014
[86] Roshani Ade, P. R. Deshmukh, An incremental ensemble of classifiers as a technique for prediction
of student's career choice, IEEE International conference on network and soft computing, July 2014
[87] Roshani Ade, P. R. Deshmukh, Classification of Students using Psychometric Tests with the help of
Incremental Naïve Bayes Algorithm, International Journal of Computer Application, Vol 89, No. 14,
pp. 27-31March,2014

Incremental learning from unbalanced data with concept class, concept drift and missing features

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Incremental learning from unbalanced data with concept class, concept drift and missing features (20)

Recently uploaded (20)

Incremental learning from unbalanced data with concept class, concept drift and missing features