New Undersampling Method Based on the kNN Approach

Class imbalance is a common problem in machine learning tasks, which often leads to sub-optimal performance of classifiers, where the classification of a new example is based on minimizing the error rate. Researchers have worked on this problem by developing various methods of resampling or modification of existing classification algorithms, however, there is still a need to look for better solutions that can overcome the limitations of known ones. In this paper, a new undersampling algorithm that is an extension of our previous proposals is proposed. Its main idea is to guarantee an even elimination of majority class objects while focusing on nearest neighbors. Extensive experiments were conducted on artificial and real datasets, and they showed that in many cases, the proposed solution outperformed other tested undersampling techniques.