Graph-based fraud detection with dual-path graph filtering

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Graph neural networks (GNNs) show promise for fraud detection due to their ability to process graph-structured data.
• Existing GNNs struggle with fraud graphs characterized by relation camouflage, high heterophily, and class imbalance.

Purpose of the Study:

• To propose a novel Graph-Based Fraud Detection Model with Dual-Path Graph Filtering (DPF-GFD) to overcome limitations of current GNNs in fraud detection.
• To enhance the discriminative and stable representation learning for nodes in complex fraud graphs.

Main Methods:

• DPF-GFD utilizes a beta wavelet-based operator for structural pattern capture and constructs a similarity graph with an improved low-pass filter.
• It fuses embeddings from original and similarity graphs via supervised representation learning.
• Node features are then used by an ensemble tree model for fraud risk assessment.

Main Results:

• The proposed DPF-GFD method demonstrates superior performance on four real-world financial fraud detection datasets.
• The dual-path filtering paradigm effectively decouples structural anomaly and feature similarity modeling.
• Achieved more discriminative and stable node representations for challenging fraud graphs.

Conclusions:

• DPF-GFD offers an effective solution for fraud detection in complex graph data, outperforming existing GNN-based approaches.
• The frequency-complementary dual-path filtering approach is a key innovation for handling heterophilous and imbalanced graph data.
• The method's effectiveness is validated through comprehensive experiments on real-world financial datasets.