Medicare Fraud Detection Based on EP-GCN

Although Graph Convolutional Networks (GCNs) have made significant progress in healthcare fraud detection, they still face challenges such as severe data imbalance and noise intentionally introduced by fraudsters. To address these issues, this study aims to solve the data imbalance problem by maxi-mizing the Area Under the ROC Curve (AUC). To mitigate the impact of noise deliberately introduced by fraudsters on the topological structure, this research enhances the performance of GCNs by incorporating a pruning algorithm. A Deep Q-Network (DQN) is employed to search for an optimal pruning strategy, reducing the influence of noisy edges on the AUC and improving the GCN's performance in node embedding and classification tasks. Experimental results demonstrate that the EP-GCN (Edge-Pruning GCN) achieves strong performance in terms of AUC and other evaluation metrics across multiple datasets.