Class imbalance presents a critical challenge in machine learning applications, where conventional classifiers often exhibit systematic bias towards the majority class, significantly impairing minority-class classification accuracy. To address this issue, this paper proposes Convex Hull Delaunay Sampling (CHDS), a novel oversampling framework that combines boundary-oriented strategies with geometric space modeling. The proposed method operates in three key phases: (1) identifying boundary-proximal minority samples using local neighborhood entropy to enhance discriminative power and mitigate noise; (2) constructing convex hulls from boundary samples to delineate distribution space of the minority class, followed by Delaunay Triangulation and safe triangle region selection within these boundaries; (3) generating synthetic samples within secure triangular regions using a Dirichlet distribution-based strategy. Comprehensive evaluations across 28 imbalanced datasets reveal that CHDS consistently outperforms 15 resampling methods in terms of AUC, F1-score, G-mean, and Recall. Ablation studies further validate the individual contributions of each core component. The framework provides a reliable data-level solution for enhancing decision-making reliability in risk-sensitive applications characterized by class imbalance.
Zheng et al. (Wed,) studied this question.