MaMiPot: a paradigm shift for the classification of imbalanced data

Abstract

The most frequently used approach for imbalance learning is resampling. In this technique, the minority class is oversampled. Most of the algorithms used for this purpose are variants of the well-known algorithm named as SMOTE, which is based on creating synthetic samples on the lines connecting selected minority instances. The variants are mainly designed to alleviate the shortcomings of SMOTE where, the idea of generating synthetic samples to avoid the bias on the majority class is preserved. This study aims at following a different path. Instead of balancing, the proposed approach is based on repositioning the samples. In particular, the classifier is enforced to learn the decision regions of the minority class by repositioning the majority class samples. Consequently, the regions in which positive instances exist will not be outnumbered the majority class samples. Hence the classifier labels these regions as the minority class. To tackle the outliers, the minority samples are repositioned in small steps to avoid distorting the original distribution. The potential of the proposed repositioning scheme is also evaluated as a preprocessing algorithm for SMOTE. The experiments that are conducted on 52 datasets from the KEEL repository have shown that the proposed approach is highly effective, when evaluated in terms of F-score, G-mean and AUC.