Reproduce “Autonomous Chain-of-Thought Distillation for Graph-Based Fraud Detection” on OpenResearch

About this paper

Yuan Li, Jun Hu, Bryan Hooi, Bingsheng He, Cheng Chen

Researchers from the National University of Singapore and ByteDance Inc. developed FraudCoT, a framework for graph-based fraud detection on text-attributed graphs that integrates autonomous Chain-of-Thought (CoT) reasoning and an efficient asymmetric co-training scheme. The system achieves state-of-the-art fraud detection performance with up to 8.8% AUPRC improvement on an industrial dataset, alongside significant training speedups and memory efficiency.

Key insights

• Autonomous, multi-hop CoT reasoning, learned through selective distillation, is critical for capturing cross-neighborhood dependencies and forming robust graph-grounded hypotheses, especially for subtle fraud patterns.
• Efficient asymmetric co-training enables practical end-to-end LLM-GNN optimization by allowing dynamic updates for target nodes while caching neighbor embeddings, thereby fostering strong semantic-structural alignment without prohibitive computational costs.
• Suppressing incorrect reasoning trajectories through negative distillation is vital for the robustness of the student LLM, preventing it from internalizing spurious correlations and making it less susceptible to misleading cues.

Summary by alphaXiv.