Anomaly Detection in Dynamic Networks through Edge-Tight Structure Embedding

Anomaly detection within dynamic networks holds significant relevance in fields such as cybersecurity, e-commerce, and social media. Recent methods employing deep learning concentrate on alterations in node-dependent substructures within dynamic networks, yielding superior outcomes compared to graph embedding-based techniques. Nevertheless, these methods grapple with two primary problems in dynamic network anomaly detection: the disruption caused by relevant neighbors during node sampling and the inadequate integration of spatio-temporal information. To address these issues, we propose a framework ESEAD for anomaly edge detection in dynamic networks in this paper. This proposed framework devises a random walk-based node sampling strategy, selecting nodes intimately associated with the edge under scrutiny, and designs an encoder model utilizing a multi-headed attention mechanism to learn the spatio-temporal features of the edge neighborhood. Subsequently, these domain-specific spatio-temporal features are fed into an anomaly discriminator for anomaly detection. Empirical results demonstrate that our proposed ESEAD framework surpasses existing state-of-the-art methods across five real-world datasets.