A weight-based approach to class imbalance and overlap problems

The class imbalance problem is one of the key challenges limiting the performance of machine learning classifiers. Existing research primarily focuses on addressing class imbalance while overlooking the significant impact of class overlap on classification performance. To address this challenge, this study proposes a weight-based approach to class imbalance and overlap problems (WCIO). In contrast to traditional undersampling strategies that simply remove samples from the overlapping regions, WCIO introduces the K-means clustering algorithm to identify these overlapping regions and establishes a sample weight assignment mechanism accordingly. Experimental validation across 20 datasets demonstrates that this method performs well in both AUC and G-mean evaluation metrics, effectively enhancing the classifier's ability to address both class imbalance and overlapping data.