Dynamic Reinforcement Learning for Suspicious Fund Flow Detection: A Multi-layer Transaction Network Approach with Adaptive Strategy Optimization

This paper proposes a dynamic reinforcement learning framework for detecting suspicious fund flows in multi-layer transaction networks. The framework integrates graph neural networks with adaptive reinforcement learning mechanisms to address the challenges of evolving money laundering patterns in financial transactions. The system architecture implements a novel multi-layer network construction approach that captures both temporal and structural characteristics of transaction patterns. A dynamic feature extraction module employs attention mechanisms and temporal convolution networks to generate comprehensive transaction representations. The reinforcement learning component utilizes a modified Deep Q-Network with prioritized experience replay to optimize detection strategies continuously. Experimental evaluation on a large-scale financial dataset comprising 10 million transactions demonstrates the framework's effectiveness. The proposed approach achieves a detection rate of 92.5% while maintaining a false positive rate below 3.68%, outperforming traditional machine learning methods and recent deep learning approaches. The framework's adaptive strategy optimization enables real-time adjustment of detection policies based on emerging patterns. Ablation studies validate the contribution of individual components, with the graph layer architecture and temporal feature extraction mechanisms showing a significant impact on system performance.