Autonomous Chain-of-Thought Distillation for Graph-Based Fraud Detection

TL;DR

FraudCoT tackles fraud detection on text-attributed graphs by enabling autonomous graph-aware reasoning through graph-grounded chain-of-thought distillation and an efficient asymmetric LLM–GNN co-training scheme. A teacher LLM generates multiple CoTs per node, which are selectively distilled into a student via positive/negative reasoning signals and integrated into node texts, producing CoT-augmented representations. The second stage trains the LLM encoder and GNN jointly in a cost-efficient manner by caching neighbor embeddings and masking full LLM calls, achieving robust semantic–structural alignment and up to 8.8% gains in AUPRC with up to 1,066x training speedups. Experimental results on public and industrial datasets demonstrate strong performance improvements and practical efficiency gains, validating the approach for real-world TAG-based fraud detection. The work advances autonomous reasoning in graph-based fraud detection and offers a scalable framework for combining reasoning with relational learning.

Abstract

Graph-based fraud detection on text-attributed graphs (TAGs) requires jointly modeling rich textual semantics and relational dependencies. However, existing LLM-enhanced GNN approaches are constrained by predefined prompting and decoupled training pipelines, limiting reasoning autonomy and weakening semantic-structural alignment. We propose FraudCoT, a unified framework that advances TAG-based fraud detection through autonomous, graph-aware chain-of-thought (CoT) reasoning and scalable LLM-GNN co-training. To address the limitations of predefined prompts, we introduce a fraud-aware selective CoT distillation mechanism that generates diverse reasoning paths and enhances semantic-structural understanding. These distilled CoTs are integrated into node texts, providing GNNs with enriched, multi-hop semantic and structural cues for fraud detection. Furthermore, we develop an efficient asymmetric co-training strategy that enables end-to-end optimization while significantly reducing the computational cost of naive joint training. Extensive experiments on public and industrial benchmarks demonstrate that FraudCoT achieves up to 8.8% AUPRC improvement over state-of-the-art methods and delivers up to 1,066x speedup in training throughput, substantially advancing both detection performance and efficiency.

Figures (11)

• Figure 1: Comparison between predefined-prompt augmentation and our graph reasoning approach.
• Figure 2: Training efficiency between naive co-training and our proposed asymmetric co-training. By performing asymmetric target and neighbor encoding, FraudCoT achieves reasonable training cost while maintaining joint optimization.
• Figure 3: Overview of the proposed FraudCoT.
• Figure 4: Ablation study on FraudCoT components.
• Figure 5: Impact of the number of distillation samples on classification performance. Teacher and Student denote the performance achieved solely by Teacher and Student LLMs, respectively.