A denoising majority weighted minority oversampling technique for imbalanced classification

The Majority Weighted Minority Oversampling Technique (MWMOTE) is a prevalent approach for addressing imbalanced classification challenges. However, MWMOTE has difficulties in identifying noise, selects only minority class instances as reference instances, and interpolates within the range of 0 to 1. To overcome the limitations, this study presents the Denoising Majority Weighted Minority Oversampling Technique (DN-MWMOTE). This innovative technique introduces an adaptive noise removal strategy that optimizes noise processing by evaluating the impact of potential noise on classification outcomes. It also expands reference selection to include both majority and minority class instances, using instance weights and k-nearest neighbor instances. Furthermore, DN-MWMOTE generates synthetic instances grounded in an instance’s significance ratio and specific linear interpolation rules. Experimental results across a synthetic dataset and 12 benchmark datasets confirm that DN-MWMOTE consistently outperforms its traditional counterparts.