Enhancing Heterogeneous Temporal Graph Neural Network with Bidirectional Multi-Scale Attention and Adaptive Loss for Real-Time Fraud Detection in Electronic Commerce

Abstract

The rapid expansion of electronic commerce has been accompanied by a corresponding rise in sophisticated transaction fraud, with global losses estimated in the hundreds of billions of dollars annually and Nigerian financial institutions alone recording ₦52.26 billion in fraud losses in 2024. Conventional fraud detection approaches, including rule-based systems and standard machine learning classifiers, treat transactions independently and therefore fail to capture the relational and temporal dependencies that characterise coordinated fraud schemes. Recent Heterogeneous Temporal Graph Neural Networks (HTGNNs) address this gap by jointly modelling spatial, temporal, and semantic relationships among transactions, cardholders, merchants, and devices, but existing formulations remain limited by unidirectional temporal encoders and loss functions that are poorly suited to severe class imbalance. This paper proposes an Enhanced Heterogeneous Temporal Graph Neural Network (E-HTGNN) that extends the baseline HTGNN architecture of Nguyen and Le (2025) through four architectural enhancements: richer behavioural feature engineering, bidirectional heterogeneous graph modelling with edge-aware message passing, a multi-scale temporal encoder combining Bidirectional LSTM and Transformer components, and an adaptive focal-loss objective for class-imbalance handling. The model was implemented on a GPU-accelerated Kaggle environment and evaluated on the PaySim synthetic mobile-money dataset and the real-world IEEE-CIS Fraud Detection dataset using AUC-ROC, PR-AUC, precision, recall, F1-score, and inference latency. On the imbalanced PaySim test set, E-HTGNN achieved an AUC-ROC of 0.9951 and PR-AUC of 0.8876, substantially outperforming the strongest reported baseline (HTGNN 2-hop: AUC-ROC 0.956, PR-AUC 0.682), while sustaining an inference latency of 1.87 ms per transaction. On the more challenging real-world IEEE-CIS dataset, the model attained an AUC-ROC of 0.841 and PR-AUC of 0.594, establishing a strong reference point for future work. These results confirm that combining bidirectional temporal encoding, entity-type-aware graph construction, and adaptive loss functions yields measurable gains in fraud detection accuracy without sacrificing real-time applicability.