A Unified Contrastive and Generative Sampling Approach for Class Imbalance Problems

Abstract

Class-imbalanced problem is common problem that needs to be addressed in machine learning especially with a class distribution where the minority class is rare, e.g. fraud detection, medical diagnosis and rare-event prediction. In this environment, traditional methods of learning tend to overfit, amplify noise and generalize poorly. To remedy these limitations, in this paper, a new class-imbalance learning approach named AHCGS (Adaptive Hybrid Contrast–Generative Sampling) is proposed. This approach combines contrastive learning with generative modeling and adpative hybrid sampling, which can adaptively reshape data distributions and increase class separability. In the process, it produces highquality synthetic examples which are very helpful for learning on the minority class. Extensive experimental results on the benchmark dataset, namely credit card fraud dataset, verify that AHCGS outperforms baseline methods in AUC-ROC and G-mean values even under extreme low false-positive conditions.