Abstract
INTRODUCTION: Graph Neural Networks (GNNs) have shown great promise in fraud detection tasks on graph-structured data. However, they struggle with challenges such as label imbalance and the presence of heterophilic neighbours, which can obscure fraudulent behaviour by embedding fraudsters among benign users. OBJECTIVES: Existing GNN-based models often address these issues by modifying the graph structure to favour homophily, frequently ignoring heterophilic neighbours during message passing. This approach weakens their capacity to detect fraud in real-world graphs. Addressing this gap is crucial to enhance the effectiveness of fraud detection in large-scale graph. METHODS: This study proposes Attention-Augmented and Depthwise Separable Convolutional Message Passing (ADSCMP), a novel GNN framework. ADSCMP actively partitions neighbours into homophilic, heterophilic, and unknown groups during message passing. It integrates lightweight attention mechanisms to highlight critical nodes and employ depthwise separable convolutions to reduce computational overhead. Furthermore, this study dynamically generates root-specific weight matrices and incorporate both spectral and spatial features to better handle complex graph topologies. RESULTS: This study evaluated ADSCMP on four benchmark datasets Amazon, YelpChi, T-Finance, and T-Social. In supervised experiments (40 % labelled data), ADSCMP achieved AUC scores of 97.91 %, 94.17 %, 97.51 %, and 99.75 %, respectively. Even with only 1 % labelled data in semi-supervised settings, the model maintained strong performance with AUCs of 93.15 %, 84.53 %, 94.92 %, and 98.59 %. Our ablation study confirmed that each model component contributed meaningfully to its overall performance. Additionally, ADSCMP reduced inference time compared to competitive baselines, making it suitable for real-time fraud detection. CONCLUSION: By actively learning from both homophilic and heterophilic neighbours and optimizing message passing with efficient convolutions and attention, ADSCMP enhances both the accuracy and scalability of fraud detection.