Target Aware Contrastive Learning for Bank Account Fraud Detection Under Extreme Imbalance and Temporal Distribution Shift

Bank account fraud detection is challenging because fraudulent applications are rare and fraud patterns may change over time. This paper proposes TTCL-FD, a Target Aware Tabular Contrastive Learning model for Bank Account Fraud Detection. The model generates additional training samples using class-conditioned feature replacement and selects informative samples based on their similarity to future data. It combines contrastive learning, class-weighted classification, and temporal feature alignment to improve fraud detection under class imbalance and distribution shift. Experiments were conducted on the Base variant of the Bank Account Fraud (BAF) dataset using two chronological evaluation splits and five random seeds. TTCL-FD achieved an average AUPRC of 0.145 and a worst-window AUPRC of 0.115, outperforming nine baseline models on both metrics. Compared with VAE + LightGBM, the strongest competing baseline in average AUPRC, TTCL-FD achieved a relative improvement of approximately 15.1%. The results suggest that TTCL-FD improves overall fraud-risk ranking and maintains stronger performance across different time periods under the evaluated settings.