A Generalizable Anomaly Detection Method in Dynamic Graphs
Xiao Yang, Xuejiao Zhao, Zhiqi Shen
TL;DR
The paper tackles anomaly detection in dynamic graphs where generalizability across tasks and datasets is challenging. It proposes GeneralDyG, a framework that combines temporal ego-graph sampling, TensGNN-based local structure extraction, and a Temporal-Aware Transformer to jointly capture local dynamics and global temporal patterns. Evaluations on four real-world datasets show that GeneralDyG outperforms state-of-the-art methods, demonstrating strong cross-task generalizability and improved efficiency via ego-graph sampling. The work offers a practical approach for robust anomaly detection in diverse dynamic graphs and points to future work on interpretability and theoretical guarantees.
Abstract
Anomaly detection aims to identify deviations from normal patterns within data. This task is particularly crucial in dynamic graphs, which are common in applications like social networks and cybersecurity, due to their evolving structures and complex relationships. Although recent deep learning-based methods have shown promising results in anomaly detection on dynamic graphs, they often lack of generalizability. In this study, we propose GeneralDyG, a method that samples temporal ego-graphs and sequentially extracts structural and temporal features to address the three key challenges in achieving generalizability: Data Diversity, Dynamic Feature Capture, and Computational Cost. Extensive experimental results demonstrate that our proposed GeneralDyG significantly outperforms state-of-the-art methods on four real-world datasets.