Learning Discriminative and Generalizable Anomaly Detector for Dynamic Graph

ICLR 2026 Conference Submission21043 Authors

Keywords: dynamic graph, anomaly detection

Abstract: Anomaly detection in dynamic graphs is critical for many real-world applications but remains challenging because labeled anomalies are scarce. Most existing approaches rely on unsupervised or semi-supervised learning, which often struggle to learn discriminative representations and generalize to unseen cases. To overcome these issues, we propose SDGAD, a supervised framework with three main components. First, we design a residual representation that highlights deviations from historical patterns, providing strong anomaly signals. Second, we constrain the residuals of normal samples within an interval defined by two co-centered hyperspheres, ensuring consistent scales while keeping anomalies separable. Third, we use a normalizing flow to model the likelihood distribution of normal samples, treating anomalies as out-of-distribution points. Based on this distribution, we derive an explicit decision boundary and further propose a bi-boundary optimization strategy to boost generalization. Experiments on six datasets, covering both real and synthetic anomalies, show that SDGAD consistently outperforms diverse baselines across multiple evaluation metrics. The code is available at this repository:\href{https://anonymous.4open.science/r/SODA-7EFD/}{https://anonymous.4open.science/r/SODA-7EFD/}.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 21043