Dynamic Task Prioritization in Meta-GNN (Graph Neural Networks) for Fraud Detection: A meta-reinforcement learning approach with adaptive graph sparsification

Detecting fraud in dynamic transaction networks presents major obstacles owing to the computational demands and the constantly changing patterns of fraudulent behavior. We propose a Dynamic Task Prioritization Meta-GNN (DTP-MetaGNN) framework merging meta-reinforcement learning (meta-RL) and adaptive graph sparsification to tackle these challenges. The framework dynamically prioritizes tasks according to emerging fraud patterns, which results in efficient meta-training and decreased computational overhead by means of edge reduction and sparse attention mechanisms. A meta-RL policy guides task sampling by weighting high-volatility fraud scenarios, where the policy’s reward function balances detection accuracy and novelty. Furthermore, the framework applies a real-time edge reduction method to trim transaction graphs while preserving only the most meaningful links. DTP-MetaGNN relies on a Temporal Graph Neural Network (T-GNN) equipped with sparse attention, which handles the sparsified graph to produce node embeddings and adjusts through meta-learning. The proposed method connects smoothly with standard transaction network components, including graph formation and attribute derivation. Principal advancements consist of merging meta-reinforcement learning-driven task prioritization with dynamic graph sparsification, executed via Proximal Policy Optimization (PPO) and Temporal Graph Attention Networks (TGAT). Experimental findings show DTP-MetaGNN attains better performance in fraud detection with a notable decrease in computational expenses. This study pushes the boundaries of current research by introducing a scalable and adaptable approach for high-speed transaction systems.