Near-Centric Synthetic Minority Over-sampling Technique for Imbalanced Dataset Learning
Xi Deng, Hongmin Ren · 2023
In the field of imbalanced learning, the challenge of imbalanced class distribution in datasets has been persistent. Over-sampling methods, especially SMOTE (Synthetic Minority Over-sampling Technique), play a crucial role in alleviating this issue. However, it still has some limitations, such as the generation of noisy samples and inadequate handling of intra-class imbalance. This paper introduces an improved over-sampling algorithm, namely NC-SMOTE (Near-Centric Synthetic Minority Over-sampling Technique), aimed at overcoming the drawbacks of SMOTE. NC-SMOTE introduces a near-centric synthesis strategy, synthesizing new samples within a safe sector to effectively avoid the generation of noisy samples and bring the new samples closer to the center of the minority class, thereby enhancing the quality of synthesized samples. Experimental results of the algorithm indicate better performance compared to SMOTE across multiple evaluation metrics, especially on the AdaBoost classifier, and it exhibits superior performance, particularly in dealing with high-dimensional datasets.