Abstract
Accurate credit card fraud detection is vital for protecting financial systems and reducing economic losses. Graph neural networks (GNNs) have shown strong potential by capturing complex patterns in transaction networks. However, existing GNN-based approaches exhibit limitations in handling class imbalance, adapting to non-graph transaction data, and capturing the relative importance of features. Therefore, we propose HMOA-GNN, a novel framework for credit card fraud detection designed to handle tabular and highly imbalanced transaction data. First, the density-driven hierarchical hybrid sampling (DEHS) module balances the dataset by generating synthetic fraudulent transactions in dense regions and removing noise. Next, the metric-optimized latent space similarity graph construction (MOLS-GC) module applies metric learning to build graphs that satisfy the homophily assumption. Finally, the Adversarially trained, feature-adaptive GraphSAGE-based model (AdaAdvSAGE) enhances feature aggregation through adversarial learning and adaptive feature selection. Experiments on multiple real-world datasets demonstrate the superior performance of our framework in credit card fraud detection.