Causal–guided dual–timescale graph neural networks for explainable fraud detection in mobile payment systems
Fraud in digital transaction and user-behavior platforms poses escalating threats to financial ecosystems, demanding detection systems that are simultaneously accurate, timely, and interpretable. Although our motivating setting is mobile payment risk control, many public fraud benchmarks contain transaction-like or behavior-fraud graphs rather than strictly mobile payment records. We therefore position the proposed framework as a general graph fraud detection model for heterogeneous interaction graphs, while explicitly analyzing the structural, temporal, and risk-pattern correspondence between payment transactions and review-spam benchmarks. Existing approaches either rely on shallow feature engineering that ignores relational structures among users, merchants, devices, and behavior entities, or employ graph-based models that treat temporal dynamics uniformly and lack mechanisms for explaining their predictions. To address these limitations, we propose TransFraudGNN, a counterfactual-guided dual-timescale graph neural network that jointly models short-term burst behavior and long-term behavioral drift through hierarchical temporal attention. A fraud-aware contrastive pretraining strategy with hard-negative mining improves representation discriminability under severe label scarcity. The explanation module identifies compact prediction-sufficient subgraphs and supports model-level counterfactual reasoning without claiming to recover the real data-generating causal mechanism. Experiments on IEEE-CIS, Amazon, and Yelp demonstrate state-of-the-art AUC-ROC values of 0.9487, 0.9361, and 0.9218, corresponding to relative AUC error reductions of 34.6%, 30.3%, and 26.1% over the strongest baseline.