Classification method for imbalanced data set based on EKCStacking algorithm

Abstract

The processing of imbalanced data sets has always been a hot issue in machine learning. The traditional classification method is to pursue the overall classification accuracy of data sets, and often ignores the classification effect of minority samples. Stacking is a framework algorithm. Based on the Stacking framework, in this paper, we introduce a new oversampling algorithm EKSMOTE and cost-sensitive theory into Stacking, and propose the EKCStacking algorithm. The algorithm uses the EKSMOTE algorithm to reduce imbalanced ratio of data set before data training, and then the Level 1 layer uses a cost-sensitive classifier. The experimental results of the data set in the Keel database show that EKCStacking improves the classification accuracy of minority samples and makes the performance more stable compared with the traditional algorithm.