MKC-SMOTE: A Novel Synthetic Oversampling Method for Multi-Class Imbalanced Data Classification

The learning of multi-class imbalance problems presents greater challenges and has fewer research results compared to binary imbalance problems. Resampling techniques are widely employed to address data imbalance problems. However, the majority of existing resampling methods are designed specifically for binary imbalance datasets and demonstrate significant limitations when applied to multi-class imbalance datasets. Therefore, this study introduces the MKC-SMOTE algorithm, a novel and effective method specifically tailored for multi-class imbalanced datasets. During the pre-processing phase, the algorithm takes into account the distribution of all classes and employs the k-nearest neighbors (kNN) algorithm to identify appropriate original samples for synthesizing minority class samples. It then utilizes an enhanced SMOTE algorithm for interpolation. In the post-processing phase, potentially misleading synthesized samples are eliminated by the undersampling technique. Consequently, the MKC-SMOTE algorithm generates high-quality minority class samples by strategically exploring the distributional regions of the classes. Extensive experiments were conducted on 21 real-world datasets, comparing the MKC-SMOTE algorithm with six imbalance problem handling methods and two classifiers. The results demonstrate that the MKC-SMOTE algorithm significantly enhances the classification performance of multi-class imbalanced datasets and outperforms several popular and state-of-the-art oversampling methods.