Geometry-Aware Contrastive Learning for Binay Imbalanced Distribution

Abstract

In binary imbalanced classification, dominant majority classes outnumber minority classes with limited samples, posing challenges for supervised contrastive learning (SupCon). Existing approaches rely on static prototypes and manual switching rules that ignore how the geometry of the representation space evolves during training, leading to suboptimal minority-class separation under severe imbalance. We address this gap with Geometry-Aware Contrastive Learning (GACL), a unified framework built on three contributions. First, we introduce learnable Equiangular Tight Frame (ETF) prototypes that are jointly optimized with the encoder, maintaining maximal angular separation throughout training rather than only at initialization. Second, we propose an uncertainty-calibrated gate that continuously interpolates between supervised and unsupervised contrastive objectives based on per-sample prediction entropy and proximity to class prototypes, removing the need for manual-tuned switching thresholds. Third, we derive a boundary-aware hardness weighting scheme that amplifies gradient contributions from genuinely difficult minority samples while attenuating those likely to be label noise, stabilizing training under extreme imbalance ratios.