Denoise-Based Over-Sampling for Imbalanced Data Classification

Abstract

Imbalanced data classification has always been a hot topic in traditional machine learning. The usual method is oversampling. Its main idea is to randomly synthesize the new minority samples between the minority samples and their neighboring samples, to put the data in a particular state of equilibrium. The existing improved methods have improved the classifier's performance to some extent, but most of the focus is on the minority sample. In this paper, a denoise-based oversampling method (DNOS) is proposed, which performs different denoise processes for the majority and minority samples. Then, it is combined with ADASYN to oversampling the data. Experimental results show that DNOS has a better classification effect than ADASYN.