Towards Anomaly Detection on Text-Attributed Graphs

ICLR 2026 Conference Submission16630 Authors

Keywords: Text attributed graph, graph anomaly detection, low-resource learning

TL;DR: The first anomaly detection framework towards Text-Attributed Graph.

Abstract: Graph anomaly detection (GAD), which aims to identify abnormal nodes that differ from the majority in graphs, has attracted considerable research attention. In real-world GAD scenarios, such as reviews in e-commerce platforms, the original features in graphs are raw text. Existing methods only treat these texts with a simple context embedding, without a comprehensive understanding of semantic information. In this work, we propose TAGAD, a novel Text-Attributed Graph Anomaly Detection framework that jointly trains the context feature and the semantic feature of texts with graph structure to detect the anomaly nodes. TAGAD consists of a global GAD module and a local GAD module, respectively for detecting global anomaly nodes and local anomaly nodes. In the global GAD module, we employ a contrastive learning strategy to jointly train the graph-text model and an autoencoder to compute the global anomaly scores. In the local GAD module, an ego graph and a text graph are constructed for each node. Then, we devise two different methods to compute local anomaly scores based on the difference between the two subgraphs, respectively for the zero-shot settings and the few-shot settings. Extensive experiments demonstrate the effectiveness of our model under both zero-shot and few-shot settings on text-attributed GAD scenarios. Codes are available at https://anonymous.4open.science/r/TAGAD-1223.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 16630