Topology-Aware Reinforcement Learning for Adaptive Credit Card Fraud Detection: Integrating Multidimensional Rebalancing with Deep Q-Networks

Abstract

Background Credit card fraud detection faces extreme class imbalance, concept drift, and topological overlap. Standard Deep Q-Networks (DQN) suffer catastrophic performance degradation (e.g., 26.8% precision) due to noisy reward signals from overlapping decision boundaries. Methods We propose the Adaptive Multidimensional Rebalancing Deep Q- Network (AMR-DQN). It integrates topology-aware data filtering directly into the DQN Experience Replay Buffer, dynamically evaluating local density, class overlap, and feature variability to weight state transitions. A cost-sensitive reward function explicitly aligns the reinforcement learning objective with institutional profitability. Results On the Kaggle Credit Card Fraud dataset, AMR-DQN achieved an F2- Score of 0.859 and AUPRC of 0.883, outperforming the strongest baseline (SMOTE- ENN + XGBoost) by 5.9% and 12.1%. It converged 43% faster than standard DQN and generated the highest Expected Fraud Loss Reduction (EFLR) of $46,750.50 per 10,000 transactions. Ablation studies revealed that removing topological constraints harms business profitability despite marginal statistical gains. Conclusions AMR-DQN bridges the gap between sequential machine learning and profit-centric financial objectives. By leveraging SHAP-based Explainable AI, it ensures regulatory compliance while providing a robust, scalable solution for adap- tive fraud detection in dynamic FinTech environments.