Cost-sensitive Fuzzy Multiple Kernel Learning for imbalanced problem

Abstract

Multiple Kernel Learning (MKL) improves the classification accuracy by exploring different formulations of data. However, its classification performance is still unsatisfactory in imbalanced problems. To address this issue, we take the characteristics of imbalanced data into account by introducing the fuzzy memberships. In our work, the fuzzy memberships are determined by both the entropies of samples and the cost for each class so as to make different samples have different contributions to the decision boundary. Thus the newly proposed method can result in more favorable classification performances on imbalanced datasets. Further, we introduce the fuzzy memberships into existing MKL to form a new algorithm Cost-sensitive Fuzzy Multiple Kernel Learning named CFMKL in short. Experimental results validate the great effectiveness of the proposed CFMKL on synthetic, real-world binary and multi-class imbalanced datasets. The major contributions of this paper are as follows. Firstly, extending the MKL to handle the imbalanced problems for the first time. Secondly, generating a new fuzzy memberships function by both the entropy and the cost-sensitive. Thirdly, proposing a new algorithm named CFMKL for imbalanced problems and validating its effectiveness.