Smart under-sampling for the detection of rare patterns in unbalanced datasets

Abstract

A novel resampling approach is presented which improves the performance of classifiers when coping with unbalanced datasets. The method selects the frequent samples, whose elimination from the training dataset is most beneficial, and automatically determines the optimal unbalance rate. The results achieved test datasets put into evidence the efficiency of themethod, that allows a sensible increase of the rare patterns detection rate and an improvement of the classification performance.