Learning frequency-aware graph fraud detection.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Graph Fraud Detection (GFD) is crucial for online systems but faces challenges like extreme label imbalance and mixed homophilic/heterophilic connections.
• Existing GFD methods often modify graph structures or suppress heterophilic neighbors, leading to bias and scalability issues.

Purpose of the Study:

• To propose a novel Frequency-aware Graph Neural Network (F-GNN) for improved GFD.
• To address limitations of existing GNNs in handling label imbalance and heterophily by adopting a frequency-domain perspective.

Main Methods:

• Developed F-GNN to decouple node representations in the graph frequency domain.
• Implemented node-adaptive spectral gating to emphasize high-frequency components.
• Introduced a fraud-aware representation fusion mechanism to mitigate label imbalance.

Main Results:

• F-GNN consistently outperformed state-of-the-art GNN-based fraud detection methods on four benchmark datasets (Yelp, Amazon, T-Finance, T-Social).
• Achieved high performance metrics, including up to 99.81% AUC and 96.65% F1-Macro in supervised and semi-supervised settings.
• Demonstrated the effectiveness of frequency-aware modeling over structure-based heuristics.

Conclusions:

• Frequency-aware modeling offers a principled approach to GFD, overcoming limitations of structure-based methods.
• F-GNN provides a robust and scalable solution for detecting fraud in complex graph data.
• The proposed method effectively handles the inherent challenges of GFD, paving the way for more reliable online systems.