A Novel Synthetic Minority Oversampling Technique for Multiclass Imbalance Problems

Abstract

Multi-class imbalanced datasets present significant challenges in many real-world classification tasks, where certain classes are severely underrepresented. This study addresses the classification problems with multi-class imbalanced datasets, which are inherently more complicated than binary imbalanced problems. To tackle this problem, a novel and effective method called the One-vs-One Center Hybrid Synthetic Minority Over-sampling Technique (OCH-SMOTE) algorithm is proposed, which combines the enhanced Synthetic Minority Oversampling Techniques (SMOTE) with the One-vs-One (OVO) decomposition strategy. The OCH-SMOTE algorithm comprises two key components: the OVO strategy is used to decompose the multi-class imbalanced datasets, and the enhanced CH-SMOTE algorithm is used to generate the balanced training datasets to improve the classification performance. The OCH-SMOTE algorithm is extensively evaluated on 18 real-world multi-class imbalanced datasets, using the CART decision tree as the base classifier. The proposed method is compared with classical and state-of-the-art oversampling methods. On average, the OCH-SMOTE algorithm improves Pmacro by 8.19%, MAVA by 9.19%, MG by 30.68%, MFM by 8.78%, and Kappa coefficient by 0.0462 across all datasets compared to the baseline methods. The experimental results demonstrate that the OCH-SMOTE algorithm significantly enhances multi-class imbalanced datasets classification performance.