E-SMOTE: Embedding-based Oversampling with Group Awareness for Fraud Detection

Abstract

Fraud detection is one of the crucial areas in today’s digital age. However, severe class imbalance often challenges it, leading to biased models struggling to detect fraud effectively. This paper proposes Embedding-based SMOTE, a subgroup-aware oversampling method based on representation learning with an Autoencoder and Contrastive Learning. To preserve subgroup structure, our method applies SMOTE within each minority subgroup (safe, borderline, rare, outlier), using increase coefficients specific to each group. Experiments show that E-SMOTE achieves competitive and balanced results across classifiers, even when more difficult samples are introduced.