Dynamic Graph Anomaly Detection via Temporal-Structural Attention and Variational Graph Autoencoder

With the increasing application of dynamic graph data in social networks, financial services and cybersecurity, dynamic graph anomaly detection has attracted growing attention. Due to the dynamicity and complexity arising from the continuous evolution of dynamic graph topology, existing methods face challenges in effectively modeling complex temporal features and capturing the intricate dependencies between temporal and structural information. These challenges hinder the comprehensive capture of spatiotemporal features in dynamic graphs, thereby limiting anomaly detection performance. To address these issues, we propose a novel dynamic graph anomaly detection method that integrates Temporal-Structural Attention graph embedding with Variational Graph Autoencoder (TSAVGA). Our approach innovatively employs a hierarchical temporal-structural attention mechanism to capture short-term spatiotemporal features, while leveraging a GCN-GRU enhanced variational graph autoencoder to model deeper long-term dependencies. In addition, we design a spectral clustering-based anomaly injection strategy to augment the diversity of anomalies in the training data. Experimental results on six real-world datasets demonstrate that TSAVGA outperforms state-of-the-art methods by 1% to 5% in AUC, and maintains robust performance across varying anomaly ratios, confirming both the effectiveness and stability of our method.