Imbalance Data Classification Based on Belief Function Theory

Jiawei Niu, Zhunga Liu

Abstract: Imbalance data is an important research for the classification and there are multiple techniques to deal with this problem. Each strategy has its particular advantage for solving imbalance data. To improve the classification performance, these strategies are combined in decision level via an appropriate way for taking fully advantages of the complementary information among different methods. Thus a new method is proposed as Evidence Redistributive Combination (ERC) for imbalance data. For query pattern, the classifier output produced by different techniques (i.e., undersampling, oversampling, hybridsampling) may have different reliabilities. So a cautious quality evaluation rule is created to estimate the credibility of each classification result based on the close neighborhoods. Then the revised classification results from different strategies are combined by Dempster’s rule to reduce the ignorant information and to generate the final classification result. Multiple experiments are used to test the performance of the new ERC method, and it shows that ERC can efficiently improve the classification performance with respect to other related methods.