Coarse to Fine: Uncovering Node Diversity in Weakly Supervised Graph Anomaly Detection

ICLR 2026 Conference Submission15307 Authors

Keywords: graph anomaly detection

Abstract: Graph anomaly detection (GAD) aims to identify abnormal nodes in graph datasets, which is a significant and challenging task. Most existing methods regard the problem as a binary classification task when exploiting the labeled data, overlooking the potential existence of fine-grained subcategories among both normal and anomalous nodes. The coarse-grained treatment often results in a sub-optimal decision boundary, and the scarcity of labeled data makes it worse. To tackle these limitations, we propose a novel framework for GAD under weak supervision, addressing the problem via two key innovations. First, we introduce a unified gating module to tackle the diversity of anomaly types. It adaptively balances node-centric attributes and neighborhood signals within a single model, allowing it to identify different anomalous patterns like contextual and structural anomalies. Second, a classifier-clustering synergy framework is developed, under which the discovery of node sub-categories and the classification of anomalies can mutually reinforce each other. We achieve this by dynamically maintaining two high-confidence sets of normal and abnormal nodes, which are determined by both of the classifier and clustering modules. Extensive experiments on seven public graph datasets demonstrate that our method consistently outperforms existing approaches, validating its effectiveness in weakly supervised graph anomaly detection.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 15307