PROBABILITY BASED CLUSTER EXPANSION OVERSAMPLING TECHNIQUE FOR IMBALANCED DATA

Dhinaharan Nagamalai et al. (Eds) : NATL, CSEA, DMDBS, Fuzzy, ITCON, NSEC, COMIT - 2018
pp. 77–90, 2018. © CS & IT-CSCP 2018 DOI : 10.5121/csit.2018.80607
PROBABILITY BASED CLUSTER
EXPANSION OVERSAMPLING TECHNIQUE
FOR IMBALANCED DATA
Shaukat Ali Shahee and Usha Ananthakumar
Shailesh J. Mehta School of Management,
Indian Institute of Technology Bombay, Mumbai, India
ABSTRACT
In many applications of data mining, class imbalance is noticed when examples in one class are
overrepresented. Traditional classifiers result in poor accuracy of the minority class due to the
class imbalance. Further, the presence of within class imbalance where classes are composed of
multiple sub-concepts with different number of examples also affect the performance of
classifier. In this paper, we propose an oversampling technique that handles between class and
within class imbalance simultaneously and also takes into consideration the generalization
ability in data space. The proposed method is based on two steps- performing Model Based
Clustering with respect to classes to identify the sub-concepts; and then computing the
separating hyperplane based on equal posterior probability between the classes. The proposed
method is tested on 10 publicly available data sets and the result shows that the proposed
method is statistically superior to other existing oversampling methods.
KEYWORDS
Supervised learning, Class Imbalance, Oversampling, Posterior Distribution
1. INTRODUCTION
Class imbalance is one of the most challenging problems in Data Mining [1]. It refers to data sets
where one class is under represented compared to another class also referred to as between class
imbalance. This phenomenon is commonly seen in many real life applications like fault detection,
fraud detection, anomaly detection, medical diagnosis [2][3][4]. Traditional classifiers applied to
such imbalanced data fail to classify minority class examples correctly due to its bias towards
majority class [5][6][7]. Owing to the large number of potential applications of class imbalance,
various methods have been proposed in the literature to address this problem. These methods can
be classified into four categories: Sampling based methods, Cost-sensitive learning, kernel-based
learning and active learning. Though various approaches exist in literature to handle class
imbalance problem, sampling based methods have shown great potential as they attempt to
improve data distribution rather than the classifier [8][9][10][11]. Liu et al. [12] have given a
number of reasons on why sampling methods are preferred compared to other methods.
Sampling based method is a pre-processing technique that diminishes the class imbalance effect
either by increasing the minority class examples or by decreasing the majority class examples

78 Computer Science & Information Technology (CS & IT)
[13][14]. In this study, we focus on oversampling as undersampling of majority class is not
recommended when the dataset has absolute rarity of minority class [15]. In case of
oversampling, the number of minority class examples is increased either by random replication of
examples or by generating new synthetic examples to minimize the overfitting problem. With
regard to synthetic sampling, the synthetic minority over-sampling technique (SMOTE) [8]
generates synthetic minority class examples. The method first randomly selects a minority class
example and then chooses its k nearest neighbours belonging to the minority class. Synthetic
examples are generated between the example under consideration and the selected nearest
neighbour example along the line joining them. However, while selecting the nearest neighbours
of the minority class, it does not consider the majority class examples and gives equal weight to
all nearest neighbours. Oversampling the examples along the line joining the considered example
and the selected nearest neighbour leads to the problem of overlapping between the classes [16].
Adaptive synthetic sampling approach for imbalanced learning (ADASYN) [17] adaptively
generates synthetic minority class examples based on their weighted distribution of minority class
examples according to the level of difficulty in learning. This method generates more synthetic
examples corresponding to hard to learn examples and less synthetic instances corresponding to
easier to learn examples. Thus the method reduces the bias due to class imbalance and adaptively
shifts the classification decision boundary towards hard to learn examples. The crux of the
method is to identify hard to learn minority class examples and ADASYN sometimes fails to find
the minority class examples that are closer to the decision boundary [9]. Majority weighted
minority oversampling technique (MWMOTE) [9] is effective in selecting hard to learn minority
class examples but in this method, small concepts present in minority class examples that are
located far from majority class examples are not identified. For handling this problem which is
also referred to as within class imbalance in literature, various cluster based methods have been
proposed in literature [18][10][19]. Cluster Based Oversampling (CBO) [19] is an oversampling
technique that can handle between-class imbalance and within-class imbalance simultaneously.
However, this method uses random oversampling to oversample the sub-clusters and thus could
result in the problem of overfitting.
Further, though class imbalance has been studied well in literature, the simultaneous presence of
between class imbalance and within class imbalance has not been addressed enough. In this
paper, we propose a method that can reasonably handle between class imbalance and within class
imbalance simultaneously. It is an oversampling approach and also considers the generalization
ability of the classifier. We have validated our proposed method on publicly available data sets
using neural network and compared with existing oversampling techniques that rely on spatial
location of minority class examples in the Euclidean feature space.
The remainder of the paper is divided into three sections. Section 2 discuses the proposed method
and its various components. Analysis on various real life data sets is presented in Section 3.
Finally, Section 4 concludes the paper with future work.
2. THE PROPOSED METHOD
The main objective of the proposed method is in enabling the classifier to give equal importance
to all the sub-clusters of the minority class that would have been otherwise lacking due to skewed
distribution of the classes. The other objective is to increase the generalization ability of the
classifier on the test dataset. Generally, the classifier tries to minimize the total error and when
the class distributions are not balanced, minimization of total error gets dominated by

Computer Science & Information Technology (CS & IT) 79
minimization of error due to majority class. Neural network is one such classifier that minimizes
the total error. The first objective is achieved by removal of between class and within class
imbalance as it helps the classifier in giving equal importance to all the sub clusters. The second
objective is realized by enlarging the data space of the sub-clusters as it increases the
generalization ability of the classifier on test set.
In the proposed method, the first step is to normalize the input dataset between [0, 1] and then to
remove the noisy examples from the dataset. A noisy example is identified based on K-Nearest
Neighbour (KNN) of the considered example. In our method, we consider an example to be noisy
if it is surrounded by 5 examples of the other class as also being considered in other studies
including [9]. Removal of noisy examples helps in reducing the oversampling of noisy examples.
After the removal of noisy examples, the concept present in data is detected using model based
clustering. The boundary of the sub-clusters is computed based on the equal posterior probability
of the classes. Subsequently, the number of examples to be oversampled is determined. Following
subsections elaborate the proposed method in detail.
2.1. Locating each sub-concept
Model based clustering is used with respect to the classes to identify the sub-clusters (or sub-
concepts) present in the dataset [20]. Model based clustering assumes that data are generated by a
mixture of probability distributions in which each component corresponds to a different cluster.
We have used MCLUST [21] for implementing the model based clustering. MCLUST is an R
package that implements the combination of hierarchical agglomerative clustering, Expectation
Maximization (EM) and the Bayesian Information criterion (BIC) for comprehensive cluster
analysis.
2.2. Locating the separating hyperplane between the sub-clusters
Model based clustering assumes that data comes from mixture of underlying probability
distributions in which each component represents a different group or cluster. In general it
considers mixture of multivariate normal distributions. In computing the separating hyperplane
between sub-clusters of two classes, the majority class sub-cluster is identified on the basis of the
nearest neighbour examples of the minority class sub-cluster. A separating hyperplane is then
computed between these two sub-clusters where the posterior probability between these two
classes are considered equal. We have
= 1| = = 0| ................................................. (1)
which is same as
| = 1 = 1 = | = 0 = 0 ................................ (2)
As oversampling handles between class and within class imbalance thus making prior probability
equal, equation (2) reduces to
| = 1 = | = 0 ............................................... (3)

Since we assume that distribution is multivariate normal, the final equation of separating
hyperplane is
− ∑ = − ∑ + /2 ....................................... (4)
where , n is the number of features of the dataset; µ1 is the mean of the minority class sub-
cluster and µ2 and ∑ are respectively the mean and covariance of the majority class sub-cluster.
After computing the separating hyperplane between the sub-clusters, we expand the size of sub-
clusters till the boundary of the region given by the hyperplane while maintaining the same
structure as shown in Figure 1.
Figure 1. Illustration of synthetic examples being generated in the enclosed region of separating hyperplane
Maintaining the same structure can be done by multiplying the data points by a scalar quantity.
The angle between the two vectors are defined as
= < , >/ | | | | .................................................. (5)
If we multiply the vectors by a scalar α
= < ! , ! >/ |! | |! | ........................................ (6)
As does not change, multiplying the vectors by a scalar does not change the structure of the
data. The scalar α is the minimum of the perpendicular distances from the centroid of the sub-
cluster to the neighbouring separating hyperplanes. After computing α, the synthetic example is
generated by randomly selecting the minority class example u and extrapolating that example by
using the following equation
" #$ℎ&$' ( )* +& = , + - − , ! ......................................(7)

In a situation where all the minority class examples of the sub-clusters lie outside the enclosing
hyperplane of the sub-cluster as shown in Figure 2, synthetic examples have been generated
inside the enclosing region of hyperplane using the following steps.
1. Let C be the centroid of the sub-cluster and vector u lie outside the region.
2. Substituting
= , + $ - − , in equation (4), we get the t value.
3. Generate the uniform number between [0-t]
4. " #$ℎ&$' ( )* +& = , + $ - − , ............................................................................ (8)
Figure 2. Illustration of synthetic examples being generated when all examples of the sub-cluster lie outside
the enclosing hyperplane
2.3. Number of examples to be oversampled
After computing the enclosed region of the sub-clusters given by hyperplane, the method then
computes the number of minority class examples to be oversampled. For this, we compute
. = . *)//0 ........................................................................ (9)
where Tsmaj is the total number of majority class examples and q is the total number of sub-
clusters of minority class. The number of examples to be oversampled in the ith
sub-cluster of
minority class is given by
$1 = . − )1 ........................................................................ (10)

where ai is the number of examples already present in the ith
sub-cluster.
2.4. Algorithm
3. COMPARATIVE ANALYSIS
In this section, we evaluate the performance of our proposed method and compare its
performance with SMOTE [10], ADASYN [12], MWMOTE [13] and CBO [16]. The proposed
method is evaluated on 10 publicly available data sets from KEEL [22] dataset repository. The
data sets considered in this study are listed in Table 1.
Input: Training dataset:
with Tsmaj = Total no of majority class example and Tsmin = Total no of minority class
examples.
Output: Oversampled Dataset
1. Remove the noisy examples from the dataset
2. Applying model based clustering with respect to the classes gives
3. A set of q minority class sub-clusters {smin1 .............sminq}
4. A set of r majority class sub-clusters {smaj1 ........... smajr }
5. For each minority class sub-cluster {smin1 .............sminq}
6. For each majority class sub-cluster {smaj1 ........... smajr }
7. Compute the separating hyperplane between the sub-cluster as explained in
section 2.2.
8. EndFor
9. EndFor
10. # Oversampling the minority class sub-clusters
11. T = Tsmaj/q
12. For each of the minority class subclusters {smin1 . . . sminq}
13. ai = size (smini)
14. ti = T - ai
15. If smini lies completely outside the enclosed region
16. then Generate synthetic examples using equation (8) of section 2.2
17. Else
18. While (ti >0)
19. S = sample (smini , ti) # Select ti examples from smini
20. Let si be the number of examples lying inside the enclosed region. Generate
synthetic examples using equation (7) as explained in section 2.2
21. ti = ti - |si|
22. EndWhile
23. EndElse
24. EndIf
25. EndFor

3.1. Data sets
As this study is about binary classification problem, we have made modifications on yeast dataset
as this is multiclass dataset, and the rest of the data sets were taken as it is. In case of yeast
dataset, it has 10 classes {MIT, NUC, CYT, ME1, ME2, ME3, EXC, VAC, POX, ERL}.We
chose ME3 as the minority class and the remaining classes were combined to form the majority
class thus making it an imbalanced dataset. Table 1 represents the characteristics of various data
sets used in this study.
Table 1. The Data sets.
Data sets Total
Examples
No. Minority
Example
No Majority
Exp
Attributes
glass1 214 76 138 9
pima 768 268 500 8
glass0 214 70 144 9
yeast1 1484 429 1055 8
vehicle2 846 218 628 18
ecoli1 336 77 259 7
yeast 1484 163 1321 8
yeast3 1484 163 1321 8
yeast-0-5-6-7-9 vs 4 528 51 477 8
yeast-0-2-5-7-9 vs 3-6-8 1004 99 905 8
3.2. Assessment metrics
Traditionally, performance of the classifier is based on accuracy and error measure that is defined
as follows
2 -3) =
.4 + .5
. $)+ ( )* +&
(33 3 3)$& = 1 − 2 -3)
where TP is the number of positive examples classified correctly and TN is the number of
negative class examples classified correctly. However, in case of imbalanced datasets, this
accuracy measure overestimates the performance of the classifier as this measure could be high
even when all or most of the minority class examples are misclassified. To deal with this
problem, Haibo [11] proposed various alternative metrics based on the confusion matrix shown in
Table 2.

Table 2. Confusion Matrix
True Class
Classifier
Output
P N
P TP FP
N FN TN
Some of the alternative measures are precision, recall, F-measure and G-mean defined as
43& ' ' # =
.4
.4 + 64
& )++ =
.4
.4 + 65
2
2
(1 ) Re Pr
Re Pr
call ecision
F Measure
cal ecision
β
β
+ × ×
− =
× +
7 − 8&)# = 9
.4
.4 + 65
∗
.5
.5 + 64
Here β is a non-negative parameter that controls the influence between precision and recall. In
this study we set β = 1 implying that both are of equal importance. Another widely used accuracy
measure is Receiving Operating Characteristic (ROC) curve that gives a graphical representation
of classifier performance. The area under this curve is known as AUC measure. Since F-measure
combines precision and recall, we provide F-measure, G-mean and AUC in this study.
3.3. Experimental setting
In this study, we use feed forward neural network with back propagation in order to evaluate the
performance of the proposed method. This particular classifier is chosen in this study as it is one
of the classifiers that minimizes the total error and the present algorithm by way of removing the
between class and within class imbalance is expected to result in improved performance of the
classifier. In the parameter setting, the number of input neurons being used is equal to the number
of features and sigmoid function is being used as the activation function with learning rate 0.3.
The number of hidden layers being used is one and the number of neurons it contains is equal to
(number of features + classes)/2 [23].
Further, in SMOTE, the number of nearest neighbours used is set to 5. In case of ADASYN, we
use k = 5 and balance level = 1. In case of MWMOTE, we set the parameters as k1 = 5, k2 = 3,
k3 = |smin|/2, cp = 3 and cf(th) = 5.

3.4. Results
The results of the 10 data sets are shown in Table 3 where in each measure, the maximum value
has been highlighted. Stratified 5-fold cross validation technique was carried out where
oversampling was carried out only in the training data containing four of the folds and the fifth
fold was used as the test set. The model was trained on the oversampled training set and applied
on the test data set in order to obtain an unbiased estimate of the model. This process was
replicated five times and its average is presented in Table 3.
Results in Table 3 show that the proposed method performs better than the other methods in most
of the data sets. It can be observed that the AUC value of the proposed method is better than other
oversampling methods except yeast1 dataset.
Table 3. F-Measure, G-Mean and AUC for 10 data sets.
Data Method F-Measure of
Majority
Class
F-Measure of
Minority
Class
G-Mean AUC
glass1 SMOTE
ADASYN
MWMOTE
CBO
Prop. Method
0.745
0.757
0.759
0.760
0.803
0.619
0.606
0.605
0.624
0.616
0.690
0.683
0.684
0.699
0.694
0.721
0.717
0.728
0.736
0.767
Pima SMOTE
ADASYN
MWMOTE
CBO
Prop. Method
0.750
0.766
0.748
0.766
0.814
0.631
0.622
0.623
0.608
0.639
0.707
0.703
0.701
0.691
0.716
0.759
0.765
0.763
0.748
0.795
glass0 SMOTE
ADASYN
MWMOTE
CBO
Prop. Method
0.817
0.819
0.812
0.797
0.842
0.681
0.680
0.678
0.661
0.713
0.762
0.761
0.758
0.743
0.785
0.832
0.820
0.819
0.795
0.847
yeast1 SMOTE
ADASYN
MWMOTE
CBO
Prop. Method
0.785
0.754
0.772
0.670
0.808
0.575
0.582
0.584
0.550
0.572
0.699
0.705
0.707
0.660
0.692
0.772
0.772
0.776
0.727
0.770
vehicle2 SMOTE
ADASYN
MWMOTE
CBO
Prop. Method
0.982
0.980
0.981
0.982
0.985
0.950
0.942
0.946
0.949
0.957
0.970
0.963
0.967
0.968
0.974
0.993
0.989
0.993
0.994
0.994

ecoli1 SMOTE
ADASYN
MWMOTE
CBO
Prop. Method
0.913
0.900
0.914
0.901
0.938
0.723
0.719
0.736
0.711
0.786
0.822
0.839
0.839
0.829
0.854
0.916
0.903
0.916
0.911
0.937
Yeast SMOTE
ADASYN
MWMOTE
CBO
Prop. Method
0.965
0.955
0.965
0.950
0.967
0.737
0.712
0.734
0.689
0.752
0.870
0.898
0.866
0.892
0.881
0.943
0.938
0.941
0.935
0.959
yeast3 SMOTE
ADASYN
MWMOTE
CBO
Prop. Method
left, ..
0.966
0.952
0.966
0.945
0.968
0.743
0.695
0.742
0.671
0.759
0.870
0.886
0.866
0.887
0.878
0.943
0.930
0.938
0.936
0.951
yeast-0-5-6-7-
9 vs 4
SMOTE
ADASYN
MWMOTE
CBO
Prop. Method
0.939
0.921
0.942
0.923
0.948
0.484
0.458
0.489
0.475
0.506
0.694
0.725
0.685
0.728
0.684
0.804
0.824
0.819
0.830
0.851
yeast-0-2-5-7-
9 vs 3-6-8
SMOTE
ADASYN
MWMOTE
CBO
Prop. Metho
0.972
0.949
0.975
0.948
0.976
0.760
0.649
0.768
0.638
0.787
0.873
0.868
0.858
0.857
0.882
0.920
0.913
0.929
0.913
0.933
To test the statistical difference between the proposed method and other existing oversampling
methods, we have performed Wilcoxon signed-rank non-parametric test [24] on the metric
measures F-measure of minority and majority class, G-mean and AUC. The null and alternative
hypotheses are as follows:
H0: The median difference is zero.
H1: The median difference is positive.
The test statistic of the Wilcoxon Signed Rank Test is defined as W = min(W+, W-) where W+ is
the sum of the positive ranks and W- is the sum of the negative ranks. As 10 data sets have been
used to carry out the test, the W value at a significance of 0.05 should be less than or equal to 10
to reject the null hypothesis. The details of Wilcoxon Signed Rank Test for AUC measure
between the proposed method and MWMOTE is given in Table 4. As we can see from this table
that W+ = 52, W- = 3, W = min(W+, W-) => W = 3, we reject the null hypothesis and conclude
that the proposed method is better than MWMOTE in terms of AUC measure.

Table 4. Wilcoxon Signed Rank Test of AUC between the proposed method and MWMOTE.
Dataset Proposed
method
MWMOTE Difference Rank
glass1 0.767 0.728 0.039 10
Pima 0.795 0.763 0.032 8.5
glass0 0.847 0.819 0.028 7
yeast1 0.770 0.776 -0.006 3
vehicle2 0.994 0.993 0.001 1
ecoli1 0.937 0.916 0.021 6
Yeast 0.959 0.941 0.018 5
yeast3 0.951 0.938 0.013 4
yeast-0-5-6-7-9 vs 4 0.851 0.819 0.032 8.5
yeast-0-2-5-7-9 vs 3-6-8 0.933 0.929 0.004 2
W+ = 52, W- = 3, W = min(52, 3) = 3
For space consideration, we present just the summary of Wilcoxon Signed Rank Test between the
proposed method and other oversampling methods for various metric measures in Table 5. From
this table, it can be seen that the proposed method is statistically significantly better than the other
oversampling methods in terms of AUC and F-measure of both majority and minority class,
although in case of G-mean, the proposed method does not seem to outperform the other
oversampling methods. Though it is desirable that any algorithm performs well on all the
measures, as stated in [25], AUC is a measure that is not sensitive to the distribution of the two
classes thus making it suitable as a performance measure for the imbalanced problem.
Table 5. Summary of Wilcoxon signed rank test between our proposed method and other methods
Method Proposed Method Metric Measure
SMOTE W+ = 55, W- = 0, W = 0
W+ = 52, W- = 3, W = 3
W+ = 36, W- = 19, W = 19
W+ = 53, W- = 2, W = 2
F-Measure of Majority class
F-Measur of Minority class
G-mean
AUC
ADASYN W+ = 55, W- = 0, W = 0
W+ = 53.5, W- = 1.5, W = 1.5
W+ = 32.5, W- = 22.5, W = 22.5
W+ = 54, W- = 1, W = 1
G-mean
AUC
MWMOTE W+ = 55, W- = 0, W = 0
W+ = 53, W- = 3, W = 3
W+ = 47.5, W- = 7.5, W = 7.5
W+ = 52, W- = 3, W = 3
G-mean
AUC
CBO W+ = 55, W- = 0, W = 0
W+ = 53.5, W- = 1.5, W = 1.5
W+ = 40, W- = 15, W = 15
W+ = 55, W- = 0, W = 0
G-mean
AUC
4. CONCLUSION
In this paper, we have proposed a method that can handle between class imbalance and within
class imbalance simultaneously. The proposed method applies model based clustering with
respect to each of the classes to identify the sub-concepts present in the dataset. Then it computes
the separating hyperplane that satisfies the equal posterior probability between the sub-concepts.

It then generates the synthetic examples while maintaining the structure of the original dataset in
the enclosed region given by the hyperplane thus increasing the generalization accuracy of the
classifier.
The proposed method has been evaluated on 10 publicly available data sets and the results clearly
show that the proposed method increases the accuracy of the classifier. However, the limitation of
the proposed method is that it gets influenced by the nearest majority class sub-clusters in the
expansion of the minority sub-clusters which could be extended as future work. Another possible
extension could be in modifying the computation of separating hyperplane by including majority
class clusters that are located far from minority class clusters.
REFERENCES
[1] Q. Yang et al., “10 Challenging Problems in Data Mining Research,” Int. J. Inf. Technol. Decis.
Mak., vol. 5, no. 4, pp. 597–604, 2006.
[2] M. Galar, A. Fernández, E. Barrenechea, and F. Herrera, “EUSBoost : Enhancing ensembles for
highly imbalanced data-sets by evolutionary undersampling,” Pattern Recognit., vol. 46, no. 12, pp.
3460–3471, 2013.
[3] S. García and F. Herrera, “Evolutionary Undersampling for Classification with Imbalanced Datasets:
Proposals and Taxonomy,” Evol. Comput., vol. 17, no. 3, pp. 275–306, 2009.
[4] S. Vajda and G. A. Fink, “Strategies for training robust neural network based digit recognizers on
unbalanced data sets,” Proc. - 12th Int. Conf. Front. Handwrit. Recognition, ICFHR 2010, no.
November 2010, pp. 148–153, 2010.
[5] S. Maldonado and J. Lopez, “Imbalanced data classification using second-order cone programming
support vector machines,” Pattern Recognit., vol. 47, no. 5, pp. 2070–2079, 2014.
[6] D. J. Yu, J. Hu, Z. M. Tang, H. Bin Shen, J. Yang, and J. Y. Yang, “Improving protein-ATP binding
residues prediction by boosting SVMs with random under-sampling,” Neurocomputing, vol. 104, pp.
180–190, 2013.
[7] C. Y. Yang, J. S. Yang, and J. J. Wang, “Margin calibration in SVM class-imbalanced learning,”
Neurocomputing, vol. 73, no. 1–3, pp. 397–411, 2009.
[8] N. V Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “SMOTE: synthetic minority over-
sampling technique,” J. Artif. Intell. Res., pp. 321–357, 2002.
[9] S. Barua, M. M. Islam, X. Yao, and K. Murase, “MWMOTE--majority weighted minority
oversampling technique for imbalanced data set learning,” Knowl. Data Eng. IEEE Trans., vol. 26,
no. 2, pp. 405–425, 2014.
[10] S.-J. Yen and Y.-S. Lee, “Cluster-based under-sampling approaches for imbalanced data
distributions,” Expert Syst. Appl., vol. 36, no. 3, pp. 5718–5727, 2009.
[11] H. Han, W. Wang, and B. Mao, “Borderline-SMOTE : A New Over-Sampling Method in,” pp. 878–
887, 2005.
[12] A. Liu, J. Ghosh, and C. E. Martin, “Generative Oversampling for Mining Imbalanced Datasets,” Int.
Conf. data Min., pp. 66–72, 2007.

[13] G. E. A. P. A. Batista, R. C. Prati, and M. C. Monard, “A Study of the Behavior of Several Methods
for Balancing Machine Learning Training Data,” SIGKDD Explor. Newsl., vol. 6, no. 1, pp. 20–29,
2004.
[14] A. Estabrooks, T. Jo, and N. Japkowicz, “A multiple resampling method for learning from
imbalanced data sets,” Comput. Intell., vol. 20, no. 1, pp. 18–36, 2004.
[15] G. M. Weiss, “Mining with Rarity: A Unifying Framework,” SIGKDD Explor., vol. 6, no. 1, pp. 7–
19, 2004.
[16] H. He and E. A. Garcia, “Learning from imbalanced data,” IEEE Trans. Knowl. Data Eng., vol. 21,
no. 9, pp. 1263–1284, 2009.
[17] H. He, Y. Bai, E. A. Garcia, and S. Li, “ADASYN: Adaptive synthetic sampling approach for
imbalanced learning,” in Proceedings of the International Joint Conference on Neural Networks,
2008, no. 3, pp. 1322–1328.
[18] C. Bunkhumpornpat, K. Sinapiromsaran, and C. Lursinsap, “DBSMOTE: Density-based synthetic
minority over-sampling technique,” Appl. Intell., vol. 36, no. 3, pp. 664–684, 2012.
[19] T. Jo and N. Japkowicz, “Class imbalances versus small disjuncts,” ACM Sigkdd Explor. Newsl., vol.
6, no. 1, pp. 40–49, 2004.
[20] C. Fraley and a E. Raftery, “Model-based clustering, discriminant analysis, and density estimation,”
J. Am. Stat. Assoc., vol. 97, no. 458, pp. 611–631, 2002.
[21] C. Fraley and A. E. Raftery, “MCLUST: Software for model-based cluster analysis,” J. Classif., vol.
16, no. 2, pp. 297–306, 1999.
[22] J. Alcalá-Fdez et al., “KEEL data-mining software tool: Data set repository, integration of algorithms
and experimental analysis framework,” J. Mult. Log. Soft Comput., vol. 17, no. 2–3, pp. 255–287,
2011.
[23] H. Guo and H. L. Viktor, “Boosting with Data Generation: Improving the Classification of Hard to
Learn Examples.,” Iea/Aie, vol. 3029, pp. 1082–1091, 2004.
[24] A. Richardson, “Nonparametric Statistics for Non-Statisticians: A Step-by-Step Approach by Gregory
W. Corder, Dale I. Foreman,” Int. Stat. Rev., vol. 78, no. 3, pp. 451–452, 2010.
[25] I. Nekooeimehr and S. K. Lai-Yuen, “Adaptive semi-unsupervised weighted oversampling (A-
SUWO) for imbalanced datasets,” Expert Syst. Appl., vol. 46, pp. 405–416, 2016.
AUTHORS
Shaukat Ali Shahee received the Bachelor's degree in Mathematics honors from Patna
University and Master's degree in computer application from the West Bengal
University of Technology, India. He is currently pursuing PhD at Shailesh J. Mehta
School of Management, Indian Institute of Technology Bombay. His research interests
include data mining, machine learning and applied statistics.

Prof. Usha Ananthakumar received PhD degree in Statistics from Indian Institute of
Technology Bombay, India. She is currently a Professor at Shailesh J. Mehta School of
Management, Indian Institute of Technology Bombay. Her research interests include
data mining, machine learning, applied Statistics and multivariate data analysis.

PROBABILITY BASED CLUSTER EXPANSION OVERSAMPLING TECHNIQUE FOR IMBALANCED DATA

More Related Content

What's hot (20)

Similar to PROBABILITY BASED CLUSTER EXPANSION OVERSAMPLING TECHNIQUE FOR IMBALANCED DATA (20)

More from cscpconf (20)

Recently uploaded (20)

PROBABILITY BASED CLUSTER EXPANSION OVERSAMPLING TECHNIQUE FOR IMBALANCED DATA