Adaptive Collaborative Minority Oversampling for Multi-Class Imbalanced Classification

Abstract

Multi-class classification tasks often encounter imbalanced data in real-world applications, misclassifying minority classes leads to severe losses. Although linear interpolation is an effective oversampling method for addressing class imbalance problem, it tends to generate noisy and overlapping examples. To tackle these challenges, we propose an adaptive collaborative minority oversampling (ACo-MO) method specifically designed for multi-class imbalanced classification. Different from existing oversampling methods that rely on k -nearest neighbors to select and generate examples, our method first leverages boosting to identify difficult-to-classify examples (e.g., those near decision boundaries and in small disjuncts). This strategy ensures that the minority class boundary is effectively expanded. Subsequently, we introduce the global distribution of the minority class and collaborate with the boosting iteration to synthesize examples, which adaptively adjusts the interpolation range to minimize noise generation. Furthermore, synthetic examples enable interpolation in the vicinity of the selected example, not just along a linear path, thereby reducing the probability of overlap among the synthesized examples. We validated the effectiveness of ACo-MO through extensive experiments on 23 datasets. The results demonstrate its superiority over ten state-of-the-art multi-class imbalanced classification methods in three performance measures.