RE-SMOTE: An Improved SMOTE Algorithm Based on Majority Class Redundancy Elimination

Class imbalance is a pervasive and challenging issue in machine learning and data mining. The traditional Synthetic Minority Over-sampling Technique (SMOTE) alleviates this problem by interpolating between minority class samples, but it often ignores the internal distribution and redundancy of the majority class, leading to severe class overlapping and blurred decision boundaries. To address these limitations, this paper proposes an improved SMOTE algorithm based on redundancy elimination, termed RE-SMOTE. The core strategy of the proposed algorithm is to eliminate redundant majority class structures before generating compensation samples. In the first stage, the spatial local density and a boundary heterogeneity ratio are defined to identify and remove core redundant samples and invasive noise from the majority class, thereby clearing and optimizing the classification boundaries. In the second stage, an adaptive weight allocation mechanism is introduced to calculate the compensation requirements for each minority class sample based on the refined boundaries, and a modified linear interpolation method is used to precisely synthesize high-quality minority samples. This "reduction before compensation" strategy effectively maintains the intrinsic topological structure of the data and clarifies the decision boundaries. The theoretical framework demonstrates that RE-SMOTE significantly reduces data redundancy, mitigates over-fitting, and enhances the classification performance of classifiers on minority classes, providing a robust solution for imbalanced data mining.