The research paper published by IJSER journal is about A Novel Class Imbalance Learning Method using Subset Filtering 1

ISSN 2229-5518

A Novel Class Imbalance Learning Method using Subset Filtering

) K. Nageswara Rao#1, Prof. T. Venkateswara rao#2 Dr. D. Rajya Lakshmi#3,

#1Research Scholar, GITAM University, Vishakhapatnam.india

#3DepartmentofComputerScienceEngineering,KLUniversity,Vijayawada,india

#3Department of Information Technology, GITAM University, Vishakhapatnam,india

nageswararaokapu@yahoo.com, , tv_venkat@yahoo.com. rdavuluri@yahoo.com

Abstract—In many real-world applications, the problem of learning from imbalanced data (the imbalanced learningproblem) is a rel atively new challenge that has attracted growing attention from both academia and industry. The imbalanced learning problem is concerned with the performance of learning algorithms in the presence of underrepresented data and severe class distribution skews. Due to the inherent complex characteristics of imbalanced data sets, learning from such data requires new understandings, principles, algorithms, and tools to transform vast amounts of raw data efficiently into information and knowledgerepresentation. In this paper, we present a new hybrid subset filtering approach for learning from skewed trainingdata. This algorithm provides a simpler and faster alternative by using C4.5 as base algorithm. We conduct experiments usingeleven UCI data sets from various application domains using f0ur base learners,and five evaluation metrics. Experimentalresults show that our method has higher Area under the ROC Curve, F-measure, precision, TP rate and TN rate val- ues than many existing class imbalance learning methods.

Index Terms— Classification, class imbalance, weighted sampling, subset filtering.

—————————  ——————————
A dataset is class imbalanced if the classification categories are not
approximately equally represented. The level of imbalance (ratio of size of the majority class to minority class) can be as huge as
1:99[1]. It is noteworthy that class imbalance is emerging as an im- portant issue in designing classifiers [2], [3], [4]. Furthermore, the class with the lowest number of instances is usually the class of in- terest from the point of view of the learning task [5]. This problem is of great interest because it turns up in many real-world classification problems, such as remote-sensing [6], pollution detection [7], risk management [8], fraud detection [9], and especially medical diagno- sis [10]–[13].

There exist techniques to develop better performing classifiers with imbalanced datasets, which are generally called Class Imbalance Learning (CIL) methods. These methods can be broadly divided into two categories, namely, external methods and internal methods. Ex- ternal methods involve preprocessing of training datasets in order to make them balanced, while internal methods deal with modifications of the learning algorithms in order to reduce their sensitiveness to class imbalance [14]. The main advantage of external methods as previously pointed out, is that they are independent of the underlying classifier. In this paper, we are laying more stress to propose an ex- ternal CIL method for solving the class imbalance problem.

This paper is organized as follows. Section 2 briefly reviews the Data Balancing problems and its measures.andin Section 3, we dis- cuss the proposed method of using the Subset filtering technique for CIL. Section 4 presents the imbalanced datasets used and measures used for validation , while In Section 5, we present the experimental setting andIn Section 6discuss, in detail, the classification results obtained by the proposed method and compare them with the results obtained by different existing methods and finally, in Section 7, we conclude the paper.

2. DATA BALANCING

Whenever a class in a classification task is underrepresented (i.e., has a lower prior probability) compared to otherclasses, we consider
the data as imbalanced [15], [16]. The main problem in imbalanced
data is that the majority classes that are represented by large numbers of patterns rule the classifier decision boundaries at the expense of the minority classes that are represented by small numbers of pat- terns. This leads to high and low accuracies in classifying the majori- ty and minority classes, respectively, which do not necessarily reflect the true difficulty in classifying these classes. Most common solu- tions to this problem balance the number of patterns in the minority or majority classes.
Either way, balancing the data has been found to alleviate the prob- lem of imbalanced data and enhance accuracy [15],[16], [17]. Data balancing is performed by, e.g., oversamplingpatterns of minority classes either randomly or from areasclose to the decision bounda- ries. Interestingly, random oversamplingis found comparable to more sophisticated oversamplingmethods [17]. Alternatively, undersam- pling isperformed on majority classes either randomly or fromareas far away from the decision boundaries. We note thatrandom under- sampling may remove significant patternsand random oversampling may lead to overfitting, sorandom sampling should be performed with care. We alsonote that, usually, oversampling of minority classes is moreaccurate than undersampling of majority classes [17].
Resampling techniques can be categorized into three groups. Under- sampling methods, which create a subset of the original data-set by eliminating instances (usually majority class instances); oversam- pling methods, which create a superset of the original data-set by replicating some instances or creating new instances from existing ones; and finally, hybrids methods that combine both sampling me- thods. Among these categories, there exist several different propos- als; from this point, we only center our attention in those that have been used in under sampling.

Random undersampling: It is a nonheuristic method that aims to balance class distribution through the random eli-

mination of majority class examples. Its major drawback is that it can discard potentially useful data, which could be important for the induction process.

The research paper published by IJSER journal is about A Novel Class Imbalance Learning Method using Subset Filtering 2

ISSN 2229-5518

Random oversampling: In the same way as random over- sampling, it tries to balance class distribution, but in this case, randomly replicating minority class instances. Several authors agree that this method can increase the likelihood of occurring overfitting, since it makes exact copies of ex- isting instances.

Hybrid Methods:In this hybrid method both undersampling and oversampling will be applied for the datasets so as to

make it a balance dataset.
The bottom line is that when studying problems with imbalanced data, using the classifiers produced by standard machine learning algorithms without adjusting the output threshold may well be a crit- ical mistake.This skewness towards minority class (positive) general- ly causes the generation of a high number of false-negative predic- tions, which lower the model’s performance on the positive class compared with the performance on the negative (majority) class.A comprehensive review of different CIL methods can be found in [18]. The following two sections briefly discuss the external- imbalance and internal-imbalance learning methods.
The external methods are independent from the learning algorithm being used, and they involve preprocessing of the training datasets to balance them before training the classifiers. Different resampling methods, such as random and focused oversampling and undersam- pling, fall into to this category. In random undersampling, the ma- jority-class examples are removed randomly, until a particular class ratio is met [19]. In random oversampling, the minority-class exam- ples are randomly duplicated, until a particular class ratio is met [18]. Synthetic minority oversamplingtechnique (SMOTE) [20] is an oversampling method, where new synthetic examples are generated in the neighborhood of the existing minority-class examples rather than directly duplicating them. In addition, several informed sam- pling methods have been introduced in [21]. A clustering-based sam- pling method has been proposed in [22], while a genetic algorithm based sampling method has been proposed in [23].

3. Class Imbalance Learning using Subset Filtering

In this section, we follow a design decomposition approach to sys-
tematically analyze the different unbalanced domains. We first brief- ly introduce the design decomposition methodology adopted for new proposed approach.

Algorithm 1 TheProposed Algorithm.

1: {Input: A set of minor class examples P, a set Ofmajorclass examples N,jPj<jN j, and T, thenumber ofsubsets to be sampled from N.}

2: i ← 0, T=N/P.

3: repeat

4: i= i+ 1

5: Randomly sample a subset NifromN,

jNij=jPj.

6: Combine P and Ni to formNPi

6: Apply filter on aNPi

7: Train and Learn A Base Classifier (C4.5)

usingNPi. Obtain the values of

AUC,TP,FP,F-Measure

7: until i= T 8: Output: Average Measure;

The different components of our proposed algorithm are elaborated in the next subsections.

3.1 Dataset Sampling

An easy way to sample a dataset is by selecting instances randomly from all classes.However, sampling in this way can break the dataset in anunequal priority way and more number of instances of the same class may be chosen in sampling. To resolve this problem and main- tain uniformity in sample, we propose a samplingstrategy called weighted component sampling.
Before creating multiple subsets, we will create the number of ma- jority subsets depending upon the number of minority instances.

3.2 Identifying number of subsets of majority class

The ratio of majority and minority instances in the unbalanced data- set is used to decide the number of subset of majority instances (T) to be created.
T= no. of majority inst(N)./no. of minority inst(P).

3.3 Applying filter

Subsets of majority instances are combined with minority subset and multiple balanced subsets are formed. Applying a specific filtering technique at this stage will help to reduce the class imbalance effects. So, Correlation based Feature Subset (CFS) filter is applied at this stage.

3.4Averaging the measures

The subsets of balanced datasets created are used to run multiple times and the resulted values are averaged to find the overall result. In results we have obtained observations for AUC, Precision, F- measure, Sensitivity, Specificity and Accuracy.

4. Datasets and measures

We considered fourbenchmark real-world imbalanceddataset from the UCI machine learning repository [24] to validateour pro- posed method. Table II summarizes the details of these datasets in the ascending order of the positive-to-negative dataset ratio. This contains the name of the dataset, the total number of examples (To- tal), attribute, the number of target classes for each dataset, number of minority class examples (#min.), the number of .majority class examples (#maj.). These datasets represent a whole variety of do- mains, complexities, and imbalance ratios.For every data set, we perform a tenfold stratified cross validation. Within each fold, the classification method is repeated ten times considering that the sam- pling of subsets introduces randomness. The AUC, Precision,
F-measure, TP rate and TN Rateof thiscross-validation process are averaged from these ten runs. The whole cross-validation process is

The research paper published by IJSER journal is about A Novel Class Imbalance Learning Method using Subset Filtering 3

ISSN 2229-5518

repeated for ten times, and the final values from this method are the averages of these ten cross-validation runs.

Evaluation Criteria:

To assess theclassification results we count the number of true posi-
tive (TP),true negative (TN), false positive (FP) (actually negative, but classifiedas positive) and false negative (FN) (actually positive, butclassified as negative) examples.It is now well known that error rate is not anappropriate evaluation criterion when there is class im- balance or unequal costs. In this paper, we use AUC, Precision, F- measure, TP Rate and TN Rate as performance evaluation measures.
To deal with class imbalance, sensitivity (or recall) and specificity have usually been adopted to monitor the classification performance on each class separately. Note that sensitivity (also called true posi- tive rate, TPrate) is the percentage of positive examplesthat are cor- rectly classified, while specificity (also referred to as true negative rate, TNrate) is defined as the proportion of negative examples that are correctly classified:
The True Positive Rate measure is computed by,
Let us define a few well known and widely used measures:

TruePositiveRate

TP

TP FN

Apart from these simple metrics, it is possible to encounter several-
The True Negative Rate measure is computed by,
more complex evaluation measures that have been used in different practical domains. One of the most popular techniques for the evalu-

TrueNegativeRate

TN

TN FP

ation of classifiers in imbalanced problems is the Receiver Operating
Characteristic (ROC) curve, which is a tool for visualizing, organiz- ing and selecting classifiers based on their tradeoffs between benefits (true positives) and costs (false positives).
A quantitative representation of a ROC curve is the area under it, which is known as AUC. When only one run is available from a clas- sifier, the AUC can be computed as the arithmetic mean (macro- average) of TPrate and TNrate:
The Area under Curve (AUC) measure is computed by,

5. Experimental Settings

A. Algorithms and Parameters

In first place, we need to define a baseline classifier which we use in our proposed algorithm implementation. With this goal, we have used C4.5 decision tree generating algorithm [25]. Furthermore, it has been widely used to deal with imbalanced data-sets [26]–[28], and C4.5 has also been included as one of the top-ten data-mining algorithms [29]. Because of these facts, we have chosen it as the most appropriate base learner. C4.5 learning algorithm constructs the decision tree top-down by the usage of the normalized information gain (difference in entropy) that results from choosing an attribute for splitting the data. The attribute with the highest normalized in- formation gain is the one used to make the decision. To validate the

AUC 1

Or

AUC

TPRATE FPRATE

2

TPRATE TN RATE

2

proposed algorithm, we compared it with the traditional C4.5,CART,REP and SMOTE. Elevenreal world benchmark data sets taken from the UCI Machine Learning Repositoryare used throughout the experiments (see Table 1). We performed theimple- mentation using Weka on Windows XP with 2Duo CPU runningon
3.16 GHz PC with 3.25 GB RAM.

2) Evaluations on Four Real-World Datasets:

On the other hand, in several problems we are especially interestedin obtaining high performance on only one class. For example, in the diagnosis of a rare disease, one of the most important things is to know how reliable a positive diagnosis is. For such problems, the precision (or purity) metric is often adopted, which
can be defined as the percentage of examples that are correctly la-
beled as positive:
The Precision measure is computed by,
We evaluate theCILSW model on four real-world datasets including Ecolic, Diabetes, Hepatitis and Breast-w datasets. The fourdatasets areobtained from the University of California at Irvine machine lear- ningrepository [24].
We then constructclassifiers from theimbalanced data based on the training dataset, and perform evaluationson the test data.We repeat this procedure ten times and use the averageof the results as the performance metric. The detailedinformation about the datasets is described in Table 1.
Pr ecision

TP

TP FP

Table 1 Summary of benchmark imbalanced datasets

Datasets # Ex.# Atts. Class (_,+)

The F-measure Value is computed by,

F measure

2 Pr ecision
Pr ecision
Re call
Re call

The research paper published by IJSER journal is about A Novel Class Imbalance Learning Method using Subset Filtering 4

ISSN 2229-5518

Diabetes 768 8 (tested-positive; tested-negative) Vote 435 16 (democrat ;republican )
Sonar 208 61 (Rock, Mine)
Sick 3772 30 (Negative, Sick)

6. Experimental Results

We have analysis the performance of our proposed algorithm on class imbalance problem in the following eleven real-world datasets. (1) Ecolic Dataset: This UCI dataset was contributed by Paul Hor- ton. Number of instances in the data set is 101, number of attributes is 7. The number of classes is 8. There are no missing values in this dataset. The results of the tenfold cross validation are shown in Table
2 From Tables 13-17, we can observe the results of Proposed
Algorithm Vs various algorithms with respect to AUC, Precision, F- measure, TP rate and TN rate. From all the tables we can conclude that Proposed algorithm has performed well in the case of AUC im- provement, Precision improvement, F-measure improvement and it is comparable in the case of TP Rate and TN Rate. The reason for the better performance of proposed algorithm is due to the multiple class nature of the dataset and the majority and minority ratio of the dataset is very low.

(2)Diabetes Dataset :The Pima Indians diabetes data set obtained from the University of California at Irvine (UCI) repository [42] contains 768 samples from two classes with 500 negative samples and 268 positive samples. The positive class is interpreted as “tested positive for diabetes.” There are eight input features for the data samples. The results of the tenfold cross validation are shown in Table 9. From Tables13- 17, we can observe the results of proposed algorithm Vs various algorithms with respect to AUC, Precision, F- measure, TP rate and TN rate. From all the tables we can conclude that proposed algorithm has given good results on AUC and tie and some underperforming results in the case of remaining measures. The Reason for the performance of proposed algorithm is the multi class nature of the dataset and the majority and minority ratio of the dataset is very high (i.e. 500:268).

(3) Hepatitis Dataset: This data set is used to diagnose whether a hepatitis patient will die or live. Number of instances in the data set is 155, number of attributes is 20, and number of classes is 2 includ- ing DIE and LIVE. There are 123 LIVE instances and 32 DIE in- stances. There are 168 missing values in this data set. The results of the tenfold cross validation are shown in Table 3. From Tables 13-17, we can observe the results of proposed algorithm Vs various algo- rithms with respect to AUC, Precision, F-measure, TP rate and TN rate. From all the tables we can conclude that proposed algorithm has given good results on all the measures. The Reason for the perfor- mance of Proposed Algorithm is the multi class nature of the dataset and the majority and minority ratio of the dataset is very high (i'e.123:32).
(4) Breast-w Dataset: This is one of the breast cancer databases at UCI, collected at the University of Wisconsin by W. H.Wolberg. The problem is to predict whether a tissue sample taken from a patient's breast is malignant or benign. There are two classes, ten numerical attributes, and 699 observations. The results of the tenfold cross va- lidation are shown in Table 7. From Tables 13-17, we can observe the results of proposed algorithm Vs various algorithms with respect to AUC, Precision, F-measure, TP rate and TN rate. From all the tables we can conclude that proposed algorithm has given moderate results on Breast-w dataset. The Reason for the performance of pro- posed algorithm is the multi class nature of the dataset and the ma- jority and minority ratio of the dataset is moderately high (i’e:
458:241).
(5) Breast Cancer Dataset: This is one of the breast cancer databases at UCI, collected at the University Medical Centre, Institute of On- cology, Ljubljana, Yugoslavia by Ming Tan and Jeff Schlimmer. There are two classes, in which 201 instances of one class and 85 instances of another class. Nine attributes, some of which are linear and some are nominal, and in total 286 observations. There are many missing values in this data set. The results of the tenfold cross valida- tion are shown in Table 6. From Tables 13-17, we can observe the results of proposed algorithm Vs various algorithms with respect to AUC, Precision, F-measure, TP rate and TN rate. From all the tables we can conclude that proposed algorithm has given moderate results on Breast-w dataset. The Reason for the performance of proposed algorithm is the multi class nature of the dataset and the majority and minority ratio of the dataset is moderately high (i’e: 201:85).
(6) Credit-g Dataset: This UCI dataset was contributed by Hans Hofmann. This UCI dataset is concerned regarding credit card appli- cations. Number of instances in the data set is 1000, number of attributes is 20, out of which 7 are numeric and 13 are nominal. The number of classes is 2. There are no missing values in this dataset. The results of the tenfold cross validation are shown in Table 8. From Tables 13-17, we can observe the results of proposed algorithm Vs various algorithms with respect to AUC, Precision, F-measure, TP rate and TN rate. From all the tables we can conclude that pro- posed algorithm has given underperformed results on Credit-g data- set. The Reason for the moderate performance of proposed algorithm is the multi class nature of the dataset and the majority and minority ratio of the dataset is moderately high (i’e: 700:300).
(7) Ionosphere Dataset: This UCI dataset was contributed by Vince Sigillito. This radar data was collected by a system in Goose Bay, Labrador. This system consists of a phased array of 16 high frequen- cy antennas with a total transmitted power on the order of 6.4 kilo- watts. Number of instances in the data set is 351, number of attributes is 34. The number of classes is 2. There are no missing attribute values. The results of the tenfold cross validation are shown in Table 4. From Tables 13-17, we can observe the results of pro- posed algorithm Vs various algorithms with respect to AUC, Preci- sion, F-measure, TP rate and TN rate. From all the tables we can conclude that proposed algorithm has given moderate results on Breast-w dataset. The Reason for the performance of proposed algo- rithm is the multi class nature of the dataset and the majority and minority ratio of the dataset is moderately high (i’e: 225:126).

The research paper published by IJSER journal is about A Novel Class Imbalance Learning Method using Subset Filtering 5

ISSN 2229-5518

Table 1. Tenfold cross validation performance for Ecolic dataset

System AUC Precision F-measure TP Rate TN Rate

Table 2.Tenfold cross validation performance for Hepatitis dataset

Table 3 Tenfold cross validation performance for ionosphere dataset

Table 4 Tenfold cross validation performance for labor dataset

____ _ _ __ _ __ _ _ __ _ __ _ _ __ _ __ ___ _ _ __ _ __ _ _ __ _ __

Table 5 Tenfold cross validation classification performance for breast_cancer dataset

Table 6.Tenfold cross validation performance for Breast_wdataset

____ _ _ __ _ __ _ _ __ _ __ _ _ __ _ __ ___ _ _ __ _ __ _ _ __ _ __ _ _ __ _

System AUC Precision F-measure TP Rate TN Rate

The research paper published by IJSER journal is about A Novel Class Imbalance Learning Method using Subset Filtering 6

ISSN 2229-5518

Table 7.Tenfold cross validation performance for Credit-g dataset

Table8.Tenfold cross validation performance for Pima Diabetes dataset

System AUC Precision F-measure TP Rate TN Rate

C4.5 0.751±0.070 0.797±0.045 0.806±0.044 0.821±0.073 0.603 ±0.111

Table 9 Tenfold cross validation performance for vote dataset

____ _ _ __ _ __ _ _ __ _ __ _ _ __ _ __ ___ _ _ __ _ __ _ _ __ _ __ _ _ __ _

Table 10 Tenfold cross validation performance for Sonar dataset

____ _ _ __ _ __ _ _ __ _ __ _ _ __ _ __ ___ _ _ __ _ __ _ _ __ _ __

Table 11 Tenfold cross validation performance for Sick dataset

The research paper published by IJSER journal is about A Novel Class Imbalance Learning Method using Subset Filtering 7

ISSN 2229-5518

Table 13. Summary of results on AUC Vs Prop. Algor.

Sick Win Win Win Win

Dataset

System C4.5 CART REP SMOTE

Table 17.Summary of results on TN Rate Vs Prop.Algor.

Ecolic Win Win Win Win Hepatitis Win Win Win Loss Ionosphere Win Win Tie Loss Labor Win Win Win Win Breast Loss Win Win Loss Breast_w Tie Win Tie Loss Credit-g Win Win Win Loss Diabetes Win Win Win Loss Vote Loss Loss Win Loss Sonar Win Win Win Loss Sick Win Win Win Win

System C4.5 CART REP SMOTE

Dataset Ecolic Win Win Win Win Hepatitis Loss Loss Loss Loss Ionosphere Loss Tie Loss Win Labor Tie Loss Win Loss Breast Win Win Win Loss Breast_w Win Win Win Loss Credit-g Win Win Win Loss Diabetes Win Win Win Loss Vote Win Win Win Win Sonar Win Loss Loss Tie Sick Win Tie Win Win

Table 14.Summary of results on Precision Vs Prop.Algor.

(8) Labor Dataset: This UCI dataset was contributed by Stan Matwin. This dataset was used to test 2tier approach with learn- ing from positive and negative examples. Number of instances in the data set is 57, number of attributes is 16, out of which 8 are numeric and 8 are nominal. The number of classes is 2. There are some missing attribute values.The results of the tenfold cross validation are shown in Table 5. From Tables 13-17, we can ob- serve the results of proposed algorithm Vs various algorithms with respect to AUC, Precision, F-measure, TP rate and TN rate. From all the tables we can conclude that proposed algorithm has given good results on Labor dataset. One the Reason for the per- formance of proposed algorithm is due to the small size, the mul- ti class nature of the dataset and the majority and minority ratio of the dataset is moderately high (i’e: 37:20).

Table 15.Summary of results on F-Measure Vs Prop.Algor.

System C4.5 CART REP SMOTE

Dataset

Table 16.Summary of results on TP RateVs Prop.Algor.

System C4.5 CART REP SMOTE

Dataset

(9) Vote Dataset: This UCI dataset was contributed by This data set is used to predict the result of a vote. It is from 1984 united states congressional voting records database. This data set in- cludes votes for each of the U.S. House of representatives con- gressmen on the 16 key votes identified by the Congressional Quarterly Almanac (CQA). The CQA lists nine different types of votes: voted for, paired for, and announced for (these three sim- plified to yea), voted against, paired against, and announced against (these three simplified to nay), voted present, voted present to avoid conflict of interest, and did not vote or otherwise make a position known (these three simplified to an unknown disposition). Number of instances is 435 (267 democrats, 168 republicans), number of attributes is 17, and number of classes is

2. There are 392 missing values in this data set. The results of the

tenfold cross validation are shown in Table 7. From Tables 13-17, we can observe the results of proposed algorithm Vs various al- gorithms with respect to AUC, Precision, F-measure, TP rate and TN rate. From all the tables we can conclude that proposed algo- rithm has given moderate results on Breast-w dataset. The Rea- son for the performance of proposed algorithm is the multi class nature of the dataset and the majority and minority ratio of the dataset is moderately high (i’e: 287:168).

(10) Sonar Dataset: This UCI dataset was contributed by Terry Sejnowski. This dataset contains 111 patterns obtained by bounc- ing sonar signals off a metal cylinder at various angles and under various conditions. This data set can be used in a number of dif-

The research paper published by IJSER journal is about A Novel Class Imbalance Learning Method using Subset Filtering 8

ISSN 2229-5518

ferent ways to test learning speed, quality of ultimate learning, ability to generalize, or combinations of these factors. Number of instances in the data set is 208, number ofattributes is 60. The number of classes is 2. There are no missing values in dataset.The results of the tenfold cross validation are shown in Table 11. From Tables 13-17, we can observe the results of proposed algo- rithm Vs various algorithms with respect to AUC, Precision, F- measure, TP rate and TN rate. From all the tables we can con- clude that proposed algorithm has given good results on Sonar dataset. One the Reason for the performance of proposed algo- rithm is due to the large size, the multi class nature of the dataset and the majority and minority ratio of the dataset is moderately high (i’e: 111:97).

(11) Sick Dataset:This UCI dataset was contributed by the Gara- van Institute and J. Ross Quinlan, New South Wales Institute, Sydney, Australia. Number of instances in the data set is 3772, number ofattributes is 29, out of which 7 are numeric and 22 are nominal. Thenumber of classes is 2. There are some missing val- ues in dataset.The results of the tenfold cross validation are shown in Table 12. From Tables 13-17, we can observe the re- sults of proposed algorithm Vs various algorithms with respect to AUC, Precision, F-measure, TP rate and TN rate. From all the tables we can conclude that proposed algorithm has given excel- lent results on Sick dataset. One the Reason for the performance of proposed algorithm is due to the very huge size of the dataset, irrelevant attributes present in the dataset, the multi class nature of the dataset and the majority and minority ratio of the dataset is moderately high (i’e: 3541:231).

7. Concusion:

In this paper we present the class imbalance problem paradigm, which exploits the subset filtering strategy in the supervised

learning research area, and implement it with C4.5 as its base learner. Experimental results show thatour proposed algorithm performed well in the case of multi class imbalance datasets. Fur- thermore, our proposed algorithm is much less volatile than C4.5. In our future work, we will apply our proposed algorithm to more learning tasks, especially high dimensional feature learning tasks.

References:

1. J. Wu, S. C. Brubaker, M. D. Mullin, and J. M. Rehg, “Fast asymmetric learning for cascade face detection,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 30, no. 3, pp. 369–382, Mar.

2008.

2. N. V. Chawla, N. Japkowicz, and A. Kotcz, Eds., Proc. ICML Workshop Learn. Imbalanced Data Sets, 2003.

3. N. Japkowicz, Ed., Proc. AAAI Workshop Learn. Imbalanced

Data Sets, 2000.\

4. G. M.Weiss, “Mining with rarity: A unifying framework,”

ACM SIGKDD Explor. Newslett., vol. 6, no. 1, pp. 7–19, Jun.

2004.

5. N. V. Chawla, N. Japkowicz, and A. Kolcz, Eds., Special Issue

Learning Imbalanced Datasets, SIGKDD Explor. Newsl.,vol.

6, no. 1, 2004.

6. W.-Z. Lu and D.Wang, “Ground-level ozone prediction by support vector machine approach with a cost-sensitive classifi-

cation scheme,” Sci. Total. Enviro., vol. 395, no. 2-3, pp. 109–

116, 2008.

7. Y.-M. Huang, C.-M. Hung, and H. C. Jiau, “Evaluation of

neural networks and data mining methods on a credit assess-

ment task for class imbalance problem,” Nonlinear Anal. R. World Appl., vol. 7, no. 4, pp. 720–747, 2006.

8. D. Cieslak, N. Chawla, and A. Striegel, “Combating imbalance in network intrusion datasets,” in IEEE Int. Conf. Granular

Comput., 2006, pp. 732–737.

9. M. A. Mazurowski, P. A. Habas, J. M. Zurada, J. Y. Lo, J. A.

Baker, and G. D. Tourassi, “Training neural network classifi-

ers for medical decision making: The effects of imbalanced da-

tasets on classification performance,” Neural Netw., vol. 21, no. 2–3, pp. 427–436, 2008.

10. A. Freitas, A. Costa-Pereira, and P. Brazdil, “Cost-sensitive

decision trees applied to medical data,” in Data Warehousing

Knowl. Discov. (Lecture Notes Series in Computer Science), I. Song, J. Eder, and T. Nguyen, Eds.,

11. K.Kilic¸,O¨ zgeUncu and I. B. Tu¨rksen, “Comparison of dif- ferent strategies of utilizing fuzzy clustering in structure identi-

fication,” Inf. Sci., vol. 177, no. 23, pp. 5153–5162, 2007.

12. M. E. Celebi, H. A. Kingravi, B. Uddin, H. Iyatomi, Y. A. As- landogan, W. V. Stoecker, and R. H. Moss, “A methodological

approach to the classification of dermoscopy images,” Com-

put.Med. Imag. Grap., vol. 31, no. 6, pp. 362–373, 2007.

13. X. Peng and I. King, “Robust BMPM training based on

second-order cone programming and its application in medical

diagnosis,” Neural Netw., vol. 21, no. 2–3, pp. 450–457,

2008.Berlin/Heidelberg, Germany: Springer, 2007, vol. 4654, pp. 303–312.

14. RukshanBatuwita and Vasile Palade (2010) FSVM-CIL: Fuzzy Support Vector Machines for Class Imbalance Learning, IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 18, NO. 3,

JUNE 2010, pp no:558-571.

15. N. Japkowicz and S. Stephen, “The Class Imbalance Problem: A Systematic Study,” Intelligent Data Analysis, vol. 6, pp.

429-450, 2002.

16. M. Kubat and S. Matwin, “Addressing the Curse of Imba- lanced Training Sets: One-Sided Selection,” Proc. 14th Int’l

Conf. Machine Learning, pp. 179-186, 1997.

17. 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 Explorations, vol. 6, pp.

20-29, 2004.1

18. D. Cieslak and N. Chawla, “Learning decision trees for unba-

lanced data,” in Machine Learning and Knowledge Discovery

in Databases. Berlin, Germany: Springer-Verlag, 2008, pp.

241–256.

19. G.Weiss, “Mining with rarity: A unifying framework,”

SIGKDD Explor.Newslett., vol. 6, no. 1, pp. 7–19, 2004.

20. N. Chawla, K. Bowyer, and P. Kegelmeyer, “SMOTE: Syn- thetic minority over-sampling technique,” J. Artif. Intell. Res.,

vol. 16, pp. 321–357, 2002.

21. J. Zhang and I. Mani, “KNN approach to unbalanced data dis-

tributions: A case study involving information extraction,” in

Proc. Int. Conf. Mach. Learning, Workshop: Learning Imba- lanced Data Sets, Washington, DC, 2003, pp. 42–48.

22. T. Jo and N. Japkowicz, “Class imbalances versus small dis-

juncts,” ACM SIGKDD Explor. Newslett., vol. 6, no. 1, pp. 40–

49, 2004.

23. S. Zou, Y. Huang, Y. Wang, J. Wang, and C. Zhou, “SVM

learning from imbalanced data by GA sampling for protein

The research paper published by IJSER journal is about A Novel Class Imbalance Learning Method using Subset Filtering 9

ISSN 2229-5518

domain prediction,” in Proc. 9th Int. Conf. Young Comput. Sci., Hunan, China, 2008, pp. 982– 987.

24. A. Asuncion D. Newman. (2007). UCI Repository of Machine

Learning Database (School of Information and Computer

Science), Irvine, CA: Univ. of California [Online]. Available:

http://www.ics.uci.edu/∼mlearn/MLRepository.htmJ. R. Quin-

lan, C4.5: Programs for Machine Learning, 1st ed. San Mateo,

CA: Morgan Kaufmann Publishers, 1993.

25. C.-T. Su and Y.-H. Hsiao, “An evaluation of the robustness of

MTS for imbalanced data,” IEEE Trans. Knowl. Data Eng.,

vol. 19, no. 10, pp. 1321– 1332, Oct. 2007.

26. [60] D. Drown, T. Khoshgoftaar, and N. Seliya, “Evolutionary

sampling and software quality modeling of high-assurance sys-

tems,” IEEE Trans. Syst., Man, Cybern. A, Syst., Humans., vol.

39, no. 5, pp. 1097–1107, Sep. 2009.

27. S. Garc´ıa, A. Fern´andez, and F. Herrera, “Enhancing the ef- fectiveness and interpretability of decision tree and rule induc-

tion classifiers with evolutionary training set selection over

imbalanced problems,” Appl. Soft Comput., vol. 9, no. 4, pp.

1304–1314, 2009.

28. X. Wu, V. Kumar, J. Ross Quinlan, J. Ghosh, Q. Yang, H. Mo- toda, G. J. McLachlan, A. Ng, B. Liu, P. S. Yu, Z.-H. Zhou,

M. Steinbach, D. J. Hand, and D. Steinberg, “Top 10 algo-

rithms in data mining,” Knowl. Inf. Syst., vol. 14, pp. 1–37,

2007.