FLARE: Structuring Chain-of-Thought Reasoning for Fraud Detection on Text-Free Graphs

摘要

Recent advancements in Large Language Models (LLMs) have sparked a surge of interest in graph learning, particularly for text-attributed graphs where nodes possess rich semantic content. However, a significant gap remains in applying LLMs to text-free graphs, such as transaction networks or web interactions, where nodes are defined by heterogeneous tabular features rather than natural language. In these scenarios, the absence of explicit semantic cues renders standard LLM-based graph approaches ineffective. To bridge this gap, we propose \texttt{FLARE}, a framework that frames fraud detection not as a classification task, but as a structured reasoning task. Instead of relying on textual attributes, we introduce a context serialization strategy that translates numerical signals, such as base detector predictions, graph topology, and dataset-level priors, into a coherent narrative for the LLM. We further design a domain-specific Chain-of-Thought mechanism that guides the LLM to act as a meta-reasoner: explicitly filtering unreliable neighbors, resolving conflicts among heterogeneous expert models, and synthesizing evidence into a final verdict. This approach effectively bridges the modality gap between text-free graph data and pre-trained language models. Empirical results on four benchmarks show that \texttt{FLARE} outperforms state-of-the-art GNNs and ensemble methods, establishing a new paradigm for LLM reasoning on non-textual structured data.