Anomaly Detection in Directed Dynamic Graphs via RDGCN and LSTAN

Anomaly detection in dynamic graphs has gained significant attention in practical applications such as cybersecurity and e-commerce. However, existing deep learning-based methods often overlook the asymmetric structural characteristics of directed dynamic graphs, limiting their applicability to such graph types. Furthermore, these methods inadequately consider the long-term and short-term temporal features of dynamic graphs, which hampers their ability to capture the evolving patterns within the graphs. This paper proposes DyGRL, an anomaly detection algorithm designed specifically for directed dynamic graphs. DyGRL utilizes a Roled-based Directed Graph Convolutional Network (RDGCN) to extract structural features from directed dynamic graphs. The RDGCN defines and aggregates node neighbor information based on their roles, effectively addressing the asymmetric nature of the graph structure. Additionally, DyGRL incorporates a Long Short-term Temporal Attention Network (LSTAN) to capture the evolution patterns of dynamic graphs. The LSTAN leverages a recurrent attention mechanism to efficiently extract and fuse both long-term and short-term temporal features, enabling a comprehensive understanding of graph evolution. We demonstrate the effectiveness and superiority of DyGRL over existing methods in detecting anomalies in directed dynamic graphs through extensive experiments on real-world datasets.