Imbalanced data resampling based on oversampling and under-sampling
Wu Le · Computer Engineering and Applications Journal · 2013
There are several aspects that might influence the performance achieved by existing learning systems in the area of machine learning. It has been reported that one of these aspects is related to class imbalance in which examples in training data belonging to one class heavily outnumber the examples in the other class. Though there are several kinds of methods to get rid of this problem, this paper only discusses using resampling method to balance data in the period of preprocessing to improve the effect of classification. There are two kinds of resampling methods:over resampling and under resampling. In this paper, four methods which combine oversampling and under-sampling method are proposed for binary classification:BSM + Tomek, BSM + ENN,CBOS+Tomek and CBOS+ENN, and present very good results for data sets with a small number of positive examples. Moreover,ten other resampling methods are also taken to make comparative experiments with the four methods proposed by this paper, and the four methods also present very good results.