A Novel Graph Fraud Detector via Grouped Attribute Completion and Confidence-Aware Contrastive Learning

Graph neural networks (GNNs) are widely used for fraud detection, but their performance often suffers in real-world scenarios due to two persistent challenges: incomplete node attributes (missing data) and extreme class imbalance (where fraud nodes are significantly rarer than benign ones). This paper introduces GFD-GC, a framework designed to address these issues simultaneously through feature completion and improved representation learning.

The authors propose a two-pronged architecture:

1. Grouped Attribute Completion: Instead of simple imputation, the model partitions a node's neighbors into three groups—benign, fraud, and unknown—based on training labels. It then performs group-specific aggregation to capture fine-grained contextual patterns, allowing it to reconstruct missing features more accurately by leveraging the heterogeneous nature of the graph.
2. Confidence-Aware Contrastive Learning: To combat the scarcity of labeled fraud nodes, the model uses a supervised contrastive learning objective. It identifies high-confidence "pseudo-fraud" nodes from the unlabeled set and includes them as anchors in the contrastive space. This forces the model to learn more compact fraud representations that are more easily separable from non-fraud nodes.

Experimental results on the Amazon and Yelp datasets show that GFD-GC consistently outperforms state-of-the-art baselines across various metrics, including Macro-F1, Recall, and AUC. The framework demonstrates significant robustness, maintaining its performance advantage even as the percentage of missing node attributes increases from 20% to 70%. Ablation studies confirm that both the grouped completion module and the contrastive learning strategy contribute positively to the final classification accuracy.

By explicitly modeling the heterogeneous neighborhood structures common in fraud graphs, this approach provides a more reliable way to handle noisy or incomplete data. The integration of pseudo-labeling into a contrastive learning framework offers a scalable way to improve detection performance in imbalanced environments where manual labeling is expensive or limited.