New Graph Neural Network Boosts Few-Shot Fraud Detection

A new research paper presents ADC-GNN, an innovative Attention-guided Diffusion-Contrastive Graph Neural Network designed to enhance fraud detection, particularly in scenarios with limited and imbalanced labeled data. This framework tackles two significant challenges in real-world fraud graphs: the scarcity of verified fraudulent labels and the issue of representation dilution, where traditional graph neural networks can either oversmooth anomalies or suppress crucial high-frequency fraud signals. ADC-GNN integrates a diffusion component for feature-space denoising augmentation, which uses contrastive learning to stabilize node representations. Additionally, a spectral attention module adaptively highlights relevant information across different hop levels and relationships within the graph. The model demonstrated consistent improvements over existing baselines on public benchmarks and a proprietary telecom dataset, especially in few-shot learning settings.