Heterophily outlier temporal aware graph neural network for online fraud detection

Online fraud detection poses increasing challenges as malicious actors continuously adapt their tactics to evade recognition. Conventional detection systems, whether rule-based or machine learning–driven, process transactions in isolation and fail to capture the interdependent, dynamic, and adversarial nature of fraud. Graph Neural Networks (GNNs) have shown promise in modeling such relational data, yet most existing models assume homophily, neglect temporal evolution, and rely heavily on supervision or graph restructuring, therefore limiting their robustness in real-world fraud scenarios characterized by heterophily, class imbalance, and camouflage. To address these issues, we propose HOT-GNN, a Heterophily Outlier Temporal-aware Graph Neural Network designed to detect complex and camouflaged fraud patterns in heterogeneous, evolving graphs. HOT-GNN introduces a Hybrid Outlier-aware Similarity (HOS) mechanism that quantifies structural, semantic, and anomaly alignment to distinguish genuine from deceptive relations. It further employs a decoupled multi-view message-passing framework to separately aggregate information from homophilic and heterophilic neighbors, preventing noise propagation and preserving discriminative heterophilic signals. Temporal positional encoding and relation-specific modules enable HOT-GNN to capture behavioral dynamics without requiring explicit dynamic graph snapshots. Extensive experiments on four benchmark datasets demonstrate that HOT-GNN achieves consistent improvements over state-of-the-art GNN-based detectors.