Risk-Aware Graph Neural Networks for Financial Fraud Detection

Financial fraud detection is critical due to expanding digital payment and increasingly sophisticated fraud. Existing Graph Neural Network (GNN) based methods face the risk signal dilution during message propagation caused by feature camouflage. To address this, we propose a Risk-Aware Graph Neural Network (RA-GNN) which explicitly models risk propagation via multi-source risk extraction and adaptive fusion, a risk-aware attention mechanism that incorporates risk scores as bias terms, a risk-gated unit that dynamically updates risk states across layers, and a masked semi-supervised training strategy that masks center node label embeddings while retaining neighbors’, forcing the model to infer risk from local graph context. Extensive experiments on three real-world datasets demonstrate that RA-GNN consistently outperforms state-of-the-art baselines across AUC, F1, and AP metrics, achieving a 13.3% relative improvement in AP on YelpChi. Ablation studies confirm the contribution of each core component.