Reliability-Based Imbalanced Data Classification with Dempster-Shafer Theory

Hongpeng Tian, Zuo-wei Zhang, Arnaud Martin, Zhunga Liu

Abstract: The classification analysis of imbalanced data remains a challenging task since the base classifier usually focuses on the majority class and ignores the minority class. This paper proposes a reliability-based imbalanced data classification approach (RIC) with Dempster-Shafer theory to address this issue. First, based on the minority class, multiple under-sampling for the majority one are implemented to obtain the corresponding balanced training sets, which results in multiple globally optimal trained classifiers. Then, the neighbors are employed to evaluate the local reliability of different classifiers in classifying each test sample, making each global optimal classifier focus on the sample locally. Finally, the revised classification results based on various local reliability are fused by the Dempster-Shafer (DS) fusion rule. Doing so, the test sample can be directly classified if more than one classifier has high local reliability. Otherwise, the neighbors belonging to different classes are employed again as the additional knowledge to revise the fusion result. The effectiveness has been verified on synthetic and several real imbalanced datasets by comparison with other related approaches.