Learning Minority Class prior to Minority Oversampling

The success of minority oversampling in dealing with class-imbalanced dataset is well manifested by existing approaches. But that do not guarantee the true class of the synthetic minority points. We address the given context in this paper, Learning Minority Class prior to Minority Oversampling (LMCMO). To guarantee the class information of synthetic minority points, we estimate the minority spaces before generating the synthetic minority points. The performance efficiency of the LMCMO oversampled dataset is tested on C4.5 decision tree and Linear Support Vector Machine (SVM) classifier. Empirical evaluations on 21 datasets using four diversified metrics indicate substantial improvement in Linear SVM outcomes of the proposed method over four competing methods. A modest but still significant gain is achieved by our method over other methods on classification using C4.5 decision tree.