InSureNet: A Graph Transformer Network Approach for Health Insurance Fraud Detection

Healthcare insurance fraud imposes a significant financial and operational burden on medical systems, necessitating robust and scalable detection methods. This paper presents InSureNet, a novel approach based on Graph Transformer Networks (GTNs) for provider-level fraud detection. Unlike conventional graph neural networks with fixed connectivity, GTNs automatically discover and learn compositional meta-paths, enabling the model to capture multi-hop, heterogeneous relationships embedded in healthcare data. By integrating meta-path discovery with graph convolution, InSureNet effectively models higher-order dependencies and structural patterns among providers. Experimental results on a publicly available healthcare fraud dataset demonstrate that InSureNet outperforms state-of-the-art graph-based methods, achieving absolute gains of 5.6 percentage points in F1-score and 17.4 percentage points in ROC-AUC. These results demonstrate the model’s effectiveness in capturing complex fraud patterns under severe class imbalance.