Research Article  Open Access
An Improved Oversampling Algorithm Based on the Samples’ Selection Strategy for Classifying Imbalanced Data
Abstract
The imbalance data refers to at least one of its classes which is usually outnumbered by the other classes. The imbalanced data sets exist widely in the real world, and the classification for them has become one of the hottest issues in the field of data mining. At present, the classification solutions for imbalanced data sets are mainly based on the algorithmlevel and the datalevel. On the datalevel, both oversampling strategies and undersampling strategies are used to realize the data balance via data reconstruction. SMOTE and RandomSMOTE are two classic oversampling algorithms, but they still possess the drawbacks such as blind interpolation and fuzzy class boundaries. In this paper, an improved oversampling algorithm based on the samples’ selection strategy for the imbalanced data classification is proposed. On the basis of the RandomSMOTE algorithm, the support vectors (SV) are extracted and are treated as the parent samples to synthesize the new examples for the minority class in order to realize the balance of the data. Lastly, the imbalanced data sets are classified with the SVM classification algorithm. Fmeasure value, Gmean value, ROC curve, and AUC value are selected as the performance evaluation indexes. Experimental results show that this improved algorithm demonstrates a good classification performance for the imbalanced data sets.
1. Introduction
1.1. Research Background
Data classification is one of the most common data mining techniques [1], whose task is to determine which target class some specific object in an unknown class belongs to. There are many classic classification algorithms of data mining [2–10]. However, each classification algorithm has its advantages and disadvantages. The effect of a classification algorithm is usually related to the characteristics of data. Among the many classification algorithms, SVM (support vector machine) classification algorithm, which is proposed by Vapnik et al. and is based on the structural risk minimization principle of the SLT, is widely used [11–16]. SVM shows many unique advantages in solving some classification problems such as small sample numbers, nonlinearity, and highdimensional pattern recognition problems. The traditional SVM classification algorithm assumes two ideas: the size of the training examples of different classes in the same data set is balanced and the cost of classification error is almost similar [2]. However, in the real world, there are many imbalanced data classification problems. In addition, the classification accuracy of the minority class is often more valuable. As mentioned above, the classification model of the SVM is based on a minimization of structural risk. Therefore, for the imbalanced data, the examples in the majority class will have a greater influence on the classifier, causing its classification weight to be in favour of the majority class and then seriously affecting the distribution of classification hyperplane. So it is very important that the classification approaches can be improved at the algorithmlevel or datalevel to solve the imbalanced data classification, which is currently a trending issue in the data mining research field.
1.2. Related Research
In the field of data classification [17–19], the classification of imbalanced data has been the focus of attention. In recent years, the classification algorithms based on imbalanced data have attracted many researchers. The international conferences on the classification of imbalanced data have been held around the world [20–22].
At present, research approaches regarding the classification of imbalanced data by the SVM mainly divide into two categories: the improving approaches at the algorithmlevel and the improving approaches at the datalevel.
At the algorithmlevel, the weighted SVM of the penalty coefficient C is used to control the different costs of misclassification error for the different classes. In general, a higher misclassification cost is imposed on the minority class, and a low misclassification cost is imposed on the majority class. In addition, AdaBoost algorithm—the integrating algorithm based on multiple single classifiers—and the improving algorithm based on kernel space are used widely.
At the datalevel, there are two main approaches: oversampling strategy for the minority samples and undersampling strategy for the majority samples.
Oversampling technique uses some approaches, such as duplicating the minority examples or artificially synthesizing new examples from the minority class with some algorithms, to balance the classes’ distribution. Chawla [23] put forward the SMOTE algorithm, which is used to synthesize the minority class examples by a linear interpolation method. The BorderlineSMOTE algorithm proposed by Han et al. [24] and the RandomSMOTE algorithm proposed by Dong [25] improved the SMOTE algorithm from different angles. Kubat and Matwin [26] discussed the mining approaches towards the samples in the imbalanced data sets. In addition, other sampling approaches [27, 28] are also classic approaches for mining minority samples.
Besides oversampling algorithm, undersampling algorithm is also a common method to deal with the imbalanced data sets. Undersampling technique mainly balances the distribution of data classes by deleting the majority class examples such as the Tomek Links algorithm [29] proposed by Wilson et al., neighbor compression rule proposed by Hart [30], and GSVMRU algorithm proposed by Tang YC et al. [31].
2. Support Vector Classification and the Optimal Classification Hyperplane
A given set includes samples: , in which . If all the examples in can be separated exactly by the hyperplane and the distance from the nearest sample point to the hyperplane is the maximum, we state that the data samples can be separated by the optimal hyperplane, which is also called the maximum margin hyperplane as shown in Figure 1.
The SVM method can also solve the classification problem of highdimensional space. We introduce the mapping and map the examples from the input space to the highdimensional feature space while the optimal classification hyperplane is constructed in the highdimensional feature space. In this way, the problem of the optimal classification hyperplane is transformed into the following optimization problem:in which is the penalty parameter, which controls the degree of penalty for misclassification samples. In addition, the greater the value of C, the greater the penalty for error.
The corresponding Lagrangian function iswhere are multipliers and . We can obtain the following dual problem:where is the kernel function.
3. Existing Algorithm Introduction
3.1. SMOTE
The oversampling technique is to increase the number of minority class samples by randomly copying the minority samples, so as to balance the examples size of the minority class and the majority class. SMOTE (synthetic minority oversampling technique) [32] is one of the most commonly used oversampling methods to solve the imbalance problems, which generates the synthetic training examples by linear interpolation for the minority class. These synthetic training examples are generated by randomly selecting one or more of the knearest neighbors for each example in the minority class. After the oversampling process, the data is reconstructed and various classification techniques can be employed for the processed data. Its basic idea is as follows.
Step 1. Setting the minority class set , for each , the knearest neighbors of are obtained by calculating the Euclidean distance between and every other sample in set .
Step 2. The sampling rate is set according to the imbalanced proportion. For each , examples are randomly selected from its knearest neighbors, and they construct the set .
Step 3. For each example , the following formula is used to generate a new example:in which represents the random number between 0 and 1.
3.2. RandomSMOTE
SMOTE is a very classic oversampling approach, but it still has some deficiencies. For example, this algorithm is prone to distribution marginalization, is unable to change the sparse distribution of the samples of the minority class [33], and interpolates blindly. In view of these problems, Dong [25] proposed an improved data oversampling algorithm, which is called RandomSMOTE algorithm. RandomSMOTE algorithm can generate new minority samples in a relatively wider region, which can effectively change the sparse distribution of the minority class samples, and largely avoiding overfitting. Its basic idea is as follows.
For each sample of the minority class set , two examples are randomly selected from the set , then this method will form a triangle region with as vertices. The process of generating new examples for the minority class is shown in Figure 2.
Step 1. Random linear interpolation is performed between and to generate temporary examples according to the sampling rate .
Step 2. Random linear interpolation is performed between and to generate new minority class examples ;
3.3. The Contribution of Different Samples to the Classifier
Kubat et al. [26] proposed the oneside selection algorithm, which divided the samples into four types: safety samples, redundant samples, boundary samples, and noise samples by judging whether some example and its knearest neighbors belong to the same category. The traditional classifier has a higher recognition for the redundant samples and the safety samples. However, the boundary samples and the noise samples, which are on the junction of the two classes, tend to be confused. These examples are called “unsafe samples,” and more attention should be paid to them for the classifier. Similarly, the SVM classification algorithm divides the samples into different types. Besides the noise and the redundant samples, SVM defines “safe samples” and “boundary samples” according to the values of Lagrange parameters [34], in which the boundary samples are called “unsafe samples.”
As is shown in Figure 3, the asterisk points represent the majority samples and the cross points represent the minority samples. The dotted circle points represent the boundary samples and the solid circle points represent the “overlapping sample points,” which belong to the noise.
3.4. A Strategy of Samples’ Selection for SVM Based on the Alien kNeighbors
For the RandomSMOTE algorithm, it is not very reasonable to choose all the minority samples to be parent samples to synthesize the new examples because only the support vectors can contribute to the classification model in the SVM classification. In this way, a large number of redundant samples in the minority class will be synthesized, and this method will inevitably lead to the quality reduction of the training samples and increase training time of SVM. Hui Han et al. believed that minority samples closed to the boundary were more likely to be misclassified, so they had a greater impact on the classification results. Therefore, the BorderlineSMOTE method [24] is proposed and the synthetic examples are only generated by those samples that are closed to the boundary according to this algorithm. Although this algorithm makes full use of valuable examples to synthesize the new examples, it still has some deficiencies. It judges whether some sample in the minority class belongs to the boundary samples only based on the number of most samples in the nearest neighbors of this sample; this is not reasonable. So this method is likely to miss some valuable examples for the classification. Therefore, the key to sampling technology lies in what methods should be adopted that can retain the distribution information of the original data in order to obtain the most critical and important samples for sampling.
Chen et al. [35] proposed a samples selection method of SVM based on the alien kneighbors. The idea of the algorithm is as follows.
The training data set is constituted via selecting the alien neighbor samples for each sample. Thus, the data set of the boundary samples is confirmed by the alien neighbors of every sample. For each sample, the algorithm can find its alien kneighbors which are close to it. Finally, this algorithm finds the union set of all the alien neighbor samples, which contain all the boundary samples that are likely to be the support vectors.
This algorithm is different from the previous knearest neighbor algorithms. For each sample, this algorithm looks for the knearest neighbors in its alien class rather than its own class. This algorithm is not restricted by boundary conditions, but still has some shortcomings: (1) because this algorithm needs to calculate the knearest neighbors for all the samples belong to a certain class, it will generate a lot of redundancy and reduce the calculation speed, and (2) for the overlapping samples on the boundaries of the positive and negative classes, they are likely to be misclassified into the data set of boundary samples and be treated as the support vectors, and it will possibly reduce the generalization ability of the classifier.
4. The AKNRandomSMOTE Algorithm
In order to avoid the deficiencies of the RandomSMOTE algorithm and the alien knearest mentioned above, we propose a new RandomSMOTE algorithm which is called AKNRandomSMOTE algorithm in this paper. The support vectors are extracted by the improved alien kneighbors algorithm, and the oversampling strategy is only performed to the boundary decision samples of the minority class rather than all the minority samples. This algorithm can reduce the computational overhead of the system and improve the classification accuracy and efficiency. Besides that, we remove the overlapping samples on the boundary of the two classes before classification, which are considered “noise samples” by us. In this way, this improved algorithm can overcome the defects of the fuzzy boundaries to some extent so that the generalization ability of the classifier can be further improved.
4.1. Deletion of the Overlapping Samples
Due to the limitations of data collection approaches, different types of data usually tend to have similar values on some attributes, which may cause the poor classification effect. These samples are located in the overlapping area of the attribute space, which are called “overlapping samples.”
The classification errors of the classifier often occur in the boundary region of the two data sets. The boundary region of the two classes is such a region where the samples overlap. The situation will become more complicated especially for the imbalanced data [36]. At the same time, aiming for fixing the fuzzy boundary problem of the RandomSMOTE algorithm, we need to judge the overlapping examples before classification and then delete them so that it can reduce the probability of fuzzy boundary of the two classes and the negative influence of the noise for classification.
Setting set , which is a set of overlapping samples; setting set , which is the minority class; , in which is the number of samples in ; setting , which is the center of the minority class , this algorithm of deleting the overlapping samples of the minority class is as follows.
Step 1. Initialize the set and the value of : , and calculate the center of the minority class .
Step 2. For each , judging whether all of the knearest neighbors of belong to the majority class, if this condition is satisfied, is added to the set .
We will repeat Step 2 until all the examples in have been searched. Then, we will judge whether is null, if , this algorithm will end; otherwise, we will turn to Step 3.
Step 3. For every sample , we will do the following:
① Calculate the distance between and the center of set , marked as ; calculate the distance between each and the center of set , marked as , in which belongs to the majority class as well as belongs to the knearest neighbors of .
② For each , if satisfying with , we will delete from ; otherwise, we will keep in .
We will repeat Step 3 until all the samples in have been searched.
Step 4. is the set of the overlapping samples of the minority class. Delete all the examples of the set .
At last, we find the overlapping samples of the majority class in the same way and delete them.
4.2. The AKNRandomSMOTE Algorithm
The sections above have introduced the drawback of the alien kneighbors algorithm, which can generate a lot of redundancy. So this algorithm must be improved. In this paper, before extracting the support vectors, we first use the Kmeans algorithm to cluster the examples for the majority class and then extract the center of each cluster. Although the number of data sets after clustering is reduced, the spatial characteristics of the data are still reserved. Therefore, we only need to look for the alien kneighbors for each cluster center. In this way, we can not only extract the support vectors but also avoid producing the redundant samples.
The steps of the AKNRandomSMOTE algorithm are as follows.
Step 1. For the samples of the majority class, the Kmeans algorithm is used for clustering. We set the number of the clusters in the majority class as . After clustering, the center of each cluster of the majority class is extracted.
Step 2. For each cluster center , we calculate its knearest neighbor samples, which belong to the minority class. All the examples constitute set , which is the set of the boundary decision samples.
Step 3. For each boundary decision sample of , the RandomSMOTE algorithm is used to generate the new samples until the number of the two classes’ samples is balanced.
Step 4. As mentioned in section above, we delete the overlapping samples for the two classes.
Step 5. The processed data is classified by SVM classification algorithm.
5. Experiments
5.1. Experimental Settings
5.1.1. Experimental Data Sets
In order to verify that this improved oversampling algorithm proposed in this paper has higher classification accuracy for the imbalanced data sets classification, we select the different algorithms for comparison, and the experiments were carried out on 10 different imbalanced data sets. In this paper, the classification accuracy of these imbalanced data sets via different processing methods is given, and these processing methods include the original classification method without oversampling processing and the above three oversampling algorithms. The classification accuracy is valued according to the values of Fmeasure and Gmean.
In this study, we used 10 imbalanced data sets which are from UCI data sets and KEEL data sets; they have different sample sizes and attributes. In addition, they are also different in class imbalance ratio (IR). Table 1 summarizes the characteristics of the imbalanced data sets selected in the experiments, including the abbreviation of the data sets names (# Abb), the number of attributes of the examples (#Attr), the number of the minority examples (# Min), the number of the majority examples (# Maj), and the imbalance ratio (IR), in which IR=#Min/#Maj.

5.1.2. Parameters’ Selection
In this paper, the AKNRandomSMOTE algorithm needs to set the following parameters: and , in which represents the number of clusters in the Kmeans algorithm, and represents the number of decision samples selected in the alien kneighbors algorithm. In order to improve the applicability of this new algorithm, we try to set the different values of and : , . In order to make the oversampling algorithms achieve their optimal classification effects and, at the same time, make the classification accuracy rate to be stable, we select the different values of the two parameters for every different data set, and they can be seen in Table 2. The value of the sampling rate is determined by the ratio of the number of the two class samples.

5.2. Performance Evaluation
The performance evaluation of the classification is an important reference for classification algorithms. The evaluation indexes, which are applied to the balanced data classification, are no longer suitable for the imbalanced data classification. The new classification performance evaluation indexes, such as precision, Recall, Fmeasure, and Gmean, are often used to measure the imbalanced classification performance [2].
Set the minority class as the positive class, and the majority class as the negative class. TP represents the number of the positive samples which are correctly classified by classifier; TN represents the number of the negative examples which are correctly classified by the classifier; FN represents the number of the positive examples which are wrongly classified by the classifier; FP represents the number of the negative examples which are wrongly classified by the classifier. The classification performance indexes for the minority class samples are as follows:
(1) Classification precision:
(2) Recall rate: the higher the value, the fewer number that the positive examples are wrongly classified by the classifier.
(3) Fmeasure: this value is a classification evaluation index which considers both recall rate and precision rate., when ; it emphasizes the effect of “precision rate”; when , it emphasizes the effect of “recall rate”; when , “precision rate” is as important as “recall rate” for the classification evaluation, which is the commonly used indicator :
(4) Gmean: the geometric average of the classification precision of the minority class and the classification precision of the majority class.Fmeasure and Gmean are two performance evaluation indexes that are usually used for the imbalanced data classification.
5.3. The Results and Analyses of the Experiments
For the imbalanced data sets mentioned above, we use the 10fold cross validation algorithm to select the training set and the test set. A subset is selected as the test set each time and the average classification accuracy of the 10times experiments is taken as a result by the cross validation algorithm.
In this paper, the SMOTE algorithm, the RandomSMOTE algorithm, and the AKNRandomSMOTE algorithm are separately used to oversample the data in order to balance the data sets, and then the SVM classification algorithm is used to realize the classification. Finally, the classification results via the original SVM classification algorithm without data processing and the classification results via the above three oversampling algorithms processing are compared.
In order to show the classification effects lively, we first classify a twodimensional imbalanced data set for the following situations: the data without processing and the data after the above three oversampling algorithms. The classification effects can be seen from Figures 4–7.
Secondly, we use the SMOTE algorithm, the RandomSMOTE algorithm, and the AKNRandomSMOTE algorithm to reconstruct the data for the 10 imbalanced data sets. Finally, we use the SVM algorithm for the classification. In Table 3, “OriginalSVM,” “SMOTE,” “RandomSMOTE,” and “AKNRandomSMOTE”, respectively, represent the SVM classification results of the imbalanced data sets without data processing and the classification results processed after the three different oversampling algorithms. We take the Fmeasure values and the Gmean values as the classification performance evaluation indicators. We calculate the values many times for each algorithm and get the stable average value. The experimental comparison results are shown in Table 3.

From Table 3, we can see the AKNRandomSMOTE algorithm proposed in this paper has higher Gmean values and Fmeasure values than other methods for the 10 data sets. For the comparison purpose, the Fmeasure average values and the Gmean average values for the data sets were further given in Figures 8 and 9 in the form of polyline graphs.
Finally, we used the statistical tools, ROC curve (Receiver Operating Characteristic Curve) and AUC value (Area under ROC Curve), to evaluate the performance of the classification methods for the imbalanced data sets. ROC curve calculates a series of sensitivity and specificity by setting a number of different critical values for the continuous variables and then takes the sensitivity as the ordinate and the specificity as the abscissa to draw a curve. The closer the ROC curve is to the upperleft coordinate point (0, 1), i.e., the larger the value of the AUC is, and the higher the classification accuracy of the classifier will be. In general, the value of AUC is between 0.5 and 1. When AUC>0.5, the closer the value of AUC is to 1, the better the classification accuracy is. We generally believe that (1) when 0.5<AUC<0.7, the classification accuracy is low; (2) when 0.7<AUC<0.9, the classifier has a certain accuracy; (3) when AUC> 0.9, the classification accuracy is higher.
Because ROC curve has a good advantage that it does not change greatly with the change of the positive and the negative samples’ distribution, ROC curve and AUC value are often used to evaluate the classification effect for the imbalanced data.
Using the data sets, Banana, Haberman, Pima, and Led as examples, we have plotted the ROC curves for the four different data processing methods (see Figures 10–13) and calculated the AUC values (see Table 4) of the ROC curves. It can be seen from the ROC curves and AUC values that the AKNRandomSMOTE algorithm proposed in this paper has better classification effect than other algorithms.

6. Conclusion
In this paper, an improved oversampling algorithm based on the samples’ selection strategy is proposed. The support vectors are extracted and treated as the parent samples for oversampling to realize the balance of the data sets. Finally, the imbalanced data sets are classified with SVM classification algorithm. The experiments compare the classification effects via the different processing methods for the 10 imbalanced data sets including the classification approach without data processing and the classification approaches with data processing by SMOTE, RandomSMOTE, and AKNRandomSMOTE. We select Fmeasure value, Gmean value, ROC curve, and AUC value as the evaluation indexes. The experimental results show that the AKNRandomSMOTE algorithm has good performance for the imbalanced data classification.
In view of the problems exiting in the experiments, two issues can be considered to be resolved in the future work. Firstly, for the Kmeans clustering algorithm and the alien kneighbors algorithm, the choices of parameters have a great impact on the experimental results. Therefore, a reasonable algorithm can be considered to optimize the parameters so that the algorithm has a stable and optimal classification effect. Secondly, the improved oversampling method can be combined with the novel or improved classification algorithm as a comprehensive solution for both the datalevel and algorithmlevel for further research.
Data Availability
The imbalanced data sets data used to support the findings of this study have been deposited in UCI data set and KEEL data set and they are available openly (the URL of the UCI data set is http://archive.ics.uci.edu/ml/datasets.php and the URL of the KEEL data set is https://sci2s.ugr.es/keel/dataset.php).
Conflicts of Interest
The authors declare that they have no conflicts of interest.
Acknowledgments
This work is supported by the National Natural Science Foundation of China 71702039.
References
 J. Han, Data Mining Concepts and Techniques, Morgan Kaufmann Publishers Inc, San Francisco, Calif, USA, 2005.
 Q. Tang, Research of Classification on Imbalanced Data Sets and Its Application in Student loans Credit Risk Management, Wuhan, China, 2012.
 R. Andrews, J. Diederich, and A. B. Tickle, “A survey and critique of techniques for extracting rules from trained artificial neural networks,” KnowledgeBased Systems, vol. 8, no. 6, pp. 373–389, 1995. View at: Publisher Site  Google Scholar
 L. Breiman, “Random forests,” Machine Learning, vol. 45, no. 1, pp. 5–32, 2001. View at: Publisher Site  Google Scholar
 P. LangIey, W. Iba, and K. Thompson, “An analysis of Bayesian classifers,” in Proceedings of the 10th National Conference of Artificial Intelligence, pp. 223–228, AAAl Press, San Jose, Calif, USA, 1992. View at: Google Scholar
 M. Ramoni and P. Sebastiani, “Robust Bayes classifiers,” Artificial Intelligence, vol. 125, no. 12, pp. 209–226, 2001. View at: Publisher Site  Google Scholar  MathSciNet
 S. Cost and S. Salzberg, “A weighted nearest neighbor algorithm for learning with symbolic features,” Machine Learning, vol. 10, no. 1, pp. 57–78, 1993. View at: Publisher Site  Google Scholar
 J. R. Quinlan, C4.5 Programs for Machine Leaning, MorganKaufmann Publishers Inc, San Francisco, Calif, USA, 1993.
 N. Cristianini and J. ShaweTaylor, An Introduction to Support Vector Machines and Other Kernelbased Learning Methods, Cambridge University Press, London, UK, 2000. View at: Publisher Site
 B. Schölkopf and A. J. Smola, Learning with Kernels: Support Vector Machines, Regularization, Optimization and Beyond, MIT Press, Cambridge, Mass, USA, 2001. View at: Publisher Site  MathSciNet
 C. J. C. Burges, “A tutorial on support vector machines for pattern recognition,” Data Mining and Knowledge Discovery, vol. 2, no. 2, pp. 121–167, 1998. View at: Publisher Site  Google Scholar
 S. Abe, Support Vector Machines for Pattern Classification, Springer, New York, NY, USA, 2005. View at: Publisher Site  MathSciNet
 H.P. Huang and Y.H. Liu, “Fuzzy support vector machines for pattern recognition and data mining,” International Journal of Fuzzy Systems, vol. 4, no. 3, pp. 826–835, 2002. View at: Google Scholar  MathSciNet
 Q. Wu, “Research on SVM learning algorithm based on optimization theory,” pp. 23, 2009. View at: Google Scholar
 B. E. Boser, I. M. Guyon, and V. N. Vapnik, “Training algorithm for optimal margin classifiers,” in Proceedings of the 5th Annual ACM Workshop on Computational Learning Theory, pp. 144–152, ACM, Pittsburgh, PA, USA, July 1992. View at: Google Scholar
 L. Qin, Research on Datamining Method for Imbalanced Dataset Based on Improved SMOTE, Nanjing university of aeronautics and astronautics, Nanjing, China, 2017.
 Y. Li et al., “The protein classification combing the site evolution distance and SVM,” Chinese Journal Computer, vol. 31, no. 1, pp. 43–50, 2008. View at: Google Scholar
 P. Xu et al., “Internet traffic classification using support vector machine,” Journal of Computer Research and Development, vol. 141, no. 3, pp. 407–414, 2009. View at: Google Scholar
 X. Yan, Comprehensive Oversampling and Underdamping Study of Imbalanced Data Sets, Northeast electric power university, Jilin City, China, 2016.
 K. Yoon and S. Kwek, “An unsupervised learning approach to resolving the data imbalanced issue in supervised learning problems in functional genomics,” in Proceedings of the Fifth International Conference on Hybrid Intelligent Systems (HIS'05), IEEE, Rio de Janeiro, Brazil, November 2005. View at: Google Scholar
 X. Chen J et al., “An application of classification analysis for skewed class distribution in therapeutic drug monitoring the case of vancomycin,” in the Proceedings of 4th International Workshop on Design of Reliable Communication Networks, pp. 35–39, IEEE, Alberta, Canada, 2003. View at: Google Scholar
 P. Radivojac, U. Korad, K. M. Sivalingam, and Z. Obradovic, “Learning from class imbalanced data in wireless sensor networks,” in Proceedings of the 2003 IEEE 58th Vehicular Technology Conference. VTC 2003Fall, pp. 3030–3034, IEEE, Orlando, FL, USA, October 2003. View at: Google Scholar
 N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “SMOTE: synthetic minority oversampling technique,” Journal of Artificial Intelligence Research, vol. 16, pp. 321–357, 2002. View at: Publisher Site  Google Scholar
 H. Han, W. Y. Wang, and B. H. Mao, “BorderlineSMOTE: a new oversampling method in imbalanced data sets learning,” in Advances in Intelligent Computing, vol. 3644 of Lecture Notes in Computer Science, pp. 878–887, Springer, Berlin, Germany, 2005. View at: Publisher Site  Google Scholar
 Y. J. Dong, The Study on RandomSMOTE for the Classification of Imbalanced Data Sets, Dalian university of technology, Dalian, China, 2009.
 M. Kubat and S. Matwin, “Addressing the curse if imbalanced training sets: one sided selection,” in Proceedings of the 14th International Conference on Machine Learning, pp. 179–186, Nashville, Tennessee, 1997. View at: Google Scholar
 R. Agarwal et al., “A new framework for learning classifier models in data mining (a casestudy in network intrusion detection),” in Proceedings of the 1st SIAM International Conference on Data Mining, pp. 1–17, Chicago, IL, USA, 2001. View at: Publisher Site  Google Scholar
 M. V. Joshi and V. Kumar, “Classification using ripple down structure (a case for rare classes),” in Proceedings of the 2004 SIAM International Conference on Data Mining, pp. 321–332, Orlando, Fla, USA, 2004. View at: Google Scholar  MathSciNet
 D. Randall Wilson and T. R. Martinez, “Reduction techniques for instancebased learning algorithms,” Machine Learning, vol. 38, no. 3, pp. 257–286, 2000. View at: Publisher Site  Google Scholar
 P. E. Hart, “The condensed nearest neighbor rule,” IEEE Transactions on Information Theory, vol. 14, no. 3, pp. 515516, 1968. View at: Publisher Site  Google Scholar
 Y. Tang, Y.Q. Zhang, and N. V. Chawla, “SVMs modeling for highly imbalanced classification,” IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics, vol. 39, no. 1, pp. 281–288, 2009. View at: Publisher Site  Google Scholar
 N. V. Chawla, A. Lazarevic, L. O. Hall, and K. W. Bowyer, “SMOTEBoost: improving prediction of the minority class in boosting,” in Proceedings of the European Conference on Principles and Practice of Knowledge Discovery in Databases, pp. 107–119, Springer, Dubrovik, Croatie, 2003. View at: Publisher Site  Google Scholar
 M. Weiss G and H. Hirsh, “A quantitative study of small disjuncts,” in Proceedings of the In Proceedings of the 17th National Conference on Artificial Intelligence, pp. 665–670, AAAI Press, Austin, TX, USA, 2000. View at: Google Scholar
 Lu. J. R., Research on classification method of highdimensional classimbalanced data sets base on SVM, Harbin Institute of Technology, Harbin, China, 2016.
 J. N. Chen et al., “Speeding up algorithm for support vector machine based on alien neighbor,” Computer Engineering, vol. 44, no. 5, pp. 19–24, 2018. View at: Google Scholar
 H. T. Xiong et al., “Study on the class overlap and its treatment methods in classification,” Journal of Management Sciences in China, vol. 4, no. 16, pp. 8–10, 2013. View at: Google Scholar
Copyright
Copyright © 2019 Wenhao Xie et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.