Several SVM Ensemble Methods Integrated with Under-Sampling for Imbalanced Data Learning

Imbalanced data learning (IDL) is one of the most active and important fields in machine learning research. This paper focuses on exploring the efficiencies of four different SVM ensemble methods integrated with under-sampling in IDL. The experimental results on 20 UCI imbalanced datasets show that two new ensemble algorithms proposed in this paper, i.e., CABagE (which is bagging-style) and MABstE (which is boosting-style), call output the SVM ensemble classifiers with better minority-class-recognition abilities than the existing ensemble methods. Further analysis on the experimental results indicates that MABstE has the best overall classification performance, and we believe that this should be attributed to its more robust example-weighting mechanism.