Temporal graph anomaly detection based on multi-view semi-supervised learning

Yifan Song

doi:10.1007/s11280-026-01429-4

Temporal graph anomaly detection based on multi-view semi-supervised learning

2026-09-17

0

PRE

AI

DOI：10.1007/s11280-026-01429-4

Original

Original request for help

Save

Abstract

En 中文

Graph anomaly detection (GAD) focuses on discovering irregular patterns that diverge markedly from typical behaviors within graph-structured data. In recent years, the integration of contrastive learning with graph neural networks (GNNs) has demonstrated considerable promise in strengthening the ability to distinguish anomalous instances. Nevertheless, the majority of current graph contrastive anomaly detection (GCAD) methods emphasize enhancing discriminative power through multi-level comparison mechanisms, yet pay limited attention to the unique complexities introduced by temporal graph settings. In real-world scenarios, temporal graphs exhibit time-evolving topological structures, complex multi-periodic dynamics, and sensitivity to unexpected events. These characteristics introduce substantial difficulty for anomaly detection, as abnormal behaviors may manifest subtly across time, emerge abruptly due to external disturbances, or evolve gradually within dynamic interaction patterns. Existing methods rarely model these temporal dependencies and multi-view structural variations in a unified framework, limiting their effectiveness in real-world dynamic environments. To overcome the aforementioned issues, we introduce Multi-View Semi-Supervised Graph Anomaly Detection (MSGAD), an innovative framework tailored for detecting anomalies in dynamic graph environments. MSGAD captures graph dynamics from multiple complementary views, enabling the model to jointly learn structural evolution patterns and temporal behavioral shifts. By integrating multi-view representations, the framework enhances robustness against noise and unexpected events while improving sensitivity to subtle anomalous signals distributed across time. Extensive evaluations on several commonly adopted benchmark datasets show that MSGAD consistently surpasses existing GAD approaches in temporal graphs. Additional ablation analyses confirm the contributions of both the multi-view modeling design and the semi-supervised learning scheme to the overall performance gains.

Keywords:

Contrastive learning

Temporal graphs

Multi-view

Semi-supervised learning

Temporal graph anomaly detection based on multi-view semi-supervised learning

Abstract

Journal

Organization

Cited Papers