Deep Temporal and Structural Embeddings for Robust Unsupervised Anomaly Detection in Dynamic Graphs

Abstract

Detecting anomalies in dynamic graphs is a complex yet essential task, as existing methods often fail to capture long-term dependencies required for identifying irregularities in evolving networks. We introduce Temporal Structural Graph Anomaly Detection (T-StructGAD), an unsupervised framework that leverages Graph Convolutional Gated Recurrent Units (GConvGRUs) and Long Short-Term Memory networks (LSTMs) to jointly model both structural and temporal dynamics in graph node embeddings. Anomalies are detected using reconstruction errors generated by an AutoEncoder, enabling the framework to robustly uncover deviations across time. Our method successfully captures temporal patterns, making it robust against subtle anomalies and structural changes. Comprehensive evaluations on four real-world datasets demonstrate that T-StructGAD consistently outperforms 12 state-of-the-art unsupervised anomaly detection models, showcasing its superior ability to detect complex anomalies in evolving graphs. This work advances anomaly detection in dynamic graphs by integrating deep learning techniques to address structural and temporal irregularities in a more effective manner.