An Improved Under-sampling Imbalanced Classification Algorithm

Abstract

To solve the problem of low classification accuracy of minority classes caused by data imbalance, an undersampling classification algorithm based on mean shift clustering for imbalanced data (UECMS) is proposed. The UECMS method uses mean shift clustering and instance selection for the samples of majority classes to complete the under-sampling. The selected samples and all the minority samples from origin data set form a new balanced data set. Also, the bagging-based ensemble learning algorithms are used to classify the balanced data sets. The experimental results show that the UECMS method improves the classification accuracy of minority classes for imbalanced data.