Beyond Influence: Decoupled Representation Learning for Dynamic Graph Anomaly Detection

ICLR 2026 Conference Submission15085 Authors

Keywords: Dynamic graph, Anomaly detection, Node influence, Flow-based generative model, Temporal networks.

Abstract:

Anomaly detection in dynamic graphs is essential for safeguarding complex systems such as social, financial, and communication networks. A fundamental challenge lies in the entanglement between node influence and anomaly signals. Node influence—measured by metrics such as PageRank—fluctuates naturally over time, yet existing methods often conflate these benign variations with anomalous behaviors, leading to false alarms or missed detections. This paper proposes DIDAN, a framework that distinguishes influence dynamics from true anomalies by separating influence-related and anomaly-related features in dynamic graphs. DIDAN integrates three components: (1) a Temporal Information Propagator that learns stable node representations by modeling local and global temporal dependencies; (2) an Anomaly Feature Synthesizer that alleviates severe class imbalance by generating diverse synthetic anomalies with a flow-based model; and (3) an Adversarial Influence-Decoupled Detector that enforces decoupling through adversarial training.
Experiments on multiple real-world dynamic graph benchmarks show that DIDAN consistently outperforms state-of-the-art methods, improving detection accuracy, robustness, and adaptability. Notably, ROC-AUC scores increased by 5.71%, 27.73%, and 1.91% on the Wikipedia, Reddit, and ALPHA datasets, respectively, highlighting the effectiveness of influence decoupling and anomaly augmentation in dynamic graph anomaly detection.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 15085