Attention-augmented and depthwise separable convolutional message passing for robust fraud detection in large-scale...

Area of Science:

• Graph Neural Networks
• Machine Learning
• Fraud Detection

Background:

• Graph Neural Networks (GNNs) show promise for fraud detection but struggle with label imbalance and heterophilic neighbors.
• Existing GNNs often modify graph structures, ignoring heterophilic connections, which limits real-world fraud detection.
• Addressing these limitations is crucial for improving large-scale graph fraud detection.

Purpose of the Study:

• Propose a novel GNN framework, Attention-Augmented and Depthwise Separable Convolutional Message Passing (ADSCMP).
• Enhance fraud detection accuracy and scalability on complex, real-world graphs.
• Improve GNN performance by effectively handling both homophilic and heterophilic neighbors.

Main Methods:

• ADSCMP partitions neighbors into homophilic, heterophilic, and unknown groups during message passing.
• Employs lightweight attention mechanisms and depthwise separable convolutions for efficiency.
• Dynamically generates root-specific weight matrices and integrates spectral and spatial features.

Main Results:

• Achieved high AUC scores on benchmark datasets (e.g., 97.91% on Amazon, 94.17% on YelpChi) in supervised settings.
• Maintained strong performance (e.g., 93.15% on Amazon, 84.53% on YelpChi) even with 1% labeled data in semi-supervised settings.
• Demonstrated reduced inference time compared to baselines, suitable for real-time applications.

Conclusions:

• ADSCMP enhances fraud detection accuracy and scalability by learning from diverse neighbor types.
• The framework's efficient message passing and attention mechanisms improve performance on complex graphs.
• ADSCMP offers a robust solution for real-time fraud detection in large-scale graph data.