LAGAD: An LLM-Augmented Framework for Graph Anomaly Detection
Abstract
Graph anomaly detection (GAD) aims to identify nodes that deviate from normal structural or semantic patterns, with broad applications in financial transactions, network security, and online social platforms. However, existing GAD methods predominantly focus on sophisticated optimization within the graph domain, such as refining neighbor selection and aggregation strategies, and thus inevitably suffer from aggregation anomaly dilution. This issue is particularly pronounced under weak-information scenarios with sparse connectivity and ambiguous features. Moreover, such methods largely overlook the rich semantic cues in raw textual content. To unleash the textual modality for compensating insufficient graph signals and mitigating aggregation anomaly dilution, large language models (LLMs) have emerged as a promising solution given their strong semantic reasoning. Nevertheless, effectively fusing textual embeddings with graph structures remains challenging. In this paper, we propose LAGAD, the first framework that introduces the LLMs-as-Consultants paradigm into GAD. LAGAD leverages LLMs to extract external knowledge from textual information to compensate for the insufficiency of graph signals, thereby enhancing GAD performance. Specifically, LAGAD identifies low-confidence nodes via dual-phase selection and builds representative two-hop subgraphs via semantic similarity. The encoded subgraphs, together with node-level textual and structural descriptions, are fed into the LLM for structure–semantic joint prediction with natural-language explanations. The LLM predictions are then converted into pseudo-labels for local retraining, forming a closed feedback loop that refines the base model. Extensive experiments demonstrate that LAGAD effectively mitigates challenges under weak-information settings and significantly enhances the performance and robustness of underlying GNN models.