Reinforcement learning with graph neural network (RL-GNN) fusion for real-time financial fraud detection: a context-aware community mining approach

The research work introduces a new framework which optimizes reinforcement learning with graph neural networks for detecting fraudulent transactions in unbalanced financial data. The proposed method connects community-swapping data mining techniques with multi-type anomaly detection algorithms. The method uses time-series patterns combined with structural properties and contextual features to detect fraudulent transactions. The model system uses Graph Attention Networks (GAT) connected to an RL controller which enhances fraud detection results through a reward mechanism that balances precision, computational efficiency, and community quality. The implemented system establishes superior results on IEEE-CIS data through an AUROC metric of 0.872 while achieving 0.683 average precision which leads to 15.7% increased discriminative power and 33% lower false positive rates when compared to GNN baseline models. The complete framework reached 0.839 F1-score, 0.872 AUROC, 0.683 average precision, and 0.54 MCC, demonstrating the beneficial effects of RL optimization beyond conventional accuracy. The model demonstrates improved detection of clustered fraud patterns, achieving a 19.7% gain in recall and a 33% reduction in false positives compared to baseline GNN models. Furthermore, the framework provides near real-time inference capability (average latency ~ 42 ms per batch) and scalability to transaction graphs with over 500K transactions, making it a practical option for deployment in financial institutions.