Three-Way Hybrid Sampling Using Granular Balls for Imbalanced Classification

Imbalanced data is prevalent in various fields, including disease diagnosis. Effective imbalanced learning methods are crucial for improving supervised learning algorithms. Existing preprocessing methods of imbalanced learning still suffer from blurred class boundaries, weak robustness, and high time costs. In this paper, a three-way hybrid sampling method using granular balls (TWHGBS) for imbalanced binary classification is proposed. First, an overlap-based granular ball generation (OGBG) method is proposed using k-division. Second, undersampling and oversampling methods are concatenated to give a new hybrid sampling method. Specifically, based on the overlap relation between granular balls (GBs), an undersampling method is proposed to identify borderline samples. An oversampling method is proposed for synthesizing minority samples within GB based on overlap degrees adaptively. Experimental results demonstrate that the TWHGBS-based classifiers exhibit superior effectiveness and robustness in terms of \(G-mean\) when compared with the existing GB-based undersampling method (GBU) and two conventional hybrid methods, namely, SMOTE-Tomek Links method (STomek) and SMOTE-ENN method (SENN).