Perturbation-Based SMOTE for Multi-Class Imbalanced Classification
Suyang Zheng, Kai Zhou, Chouyong Chen · 2024
Multi-class imbalanced classification is a challenge in machine learning. The classical synthetic minority oversampling technique (SMOTE) alleviates this challenge by balancing the class distribution. However, it tends to generate many noisy examples and synthesize some invalid examples. To address these issues, we propose a perturbation-based SMOTE method for multi-class imbalanced classification, which focuses on example perception in boundary regions. Different from the existing methods of randomly select anchor examples, we first consider the global class distribution, and select the examples that are difficult to classify near class boundaries as anchor examples. This ensures that synthetic examples are within class boundaries, avoiding invalid interpolation. Then, the nearest neighbors are used to perturb the synthetic examples in the feature space, which avoids overlap between examples and increases the diversity of minority classes. Finally, iterative-partitioning filter is introduced to delete the existing noisy examples. Experiments with real datasets demonstrate that our proposed method achieves competitive performance.