Hypergraph-Driven Anomaly Detection in Dynamic Noisy Graphs

As interactions among elements in applications such as social networks, transaction networks, and IP-IP networks dynamically evolve, anomaly detection in dynamic graphs to mitigate potentially threatening interactions has become increasingly important. Existing methods often assume noise-free graph structures and primarily focus on monitoring structural changes to discover anomalies. Regrettably, practical applications often involve inaccurate information, individual non-response and dropout, and sampling biases. These factors contribute to the pervasiveness of dynamic noisy graphs that encompass structural noises, making anomaly detection more challenging. To address this issue, we propose a novel Hypergraph-driven Anomaly Detection Framework (HADF), which resists the interference of structural noises and adapts to dynamic noisy graphs. HADF consists of a hyper encoder and an embedding enhancer. The hyper encoder leverages inter-edge correlations to generate hyperedges and design their resistant weights, further employing hypergraph convolutional layers to extract the basic hyper-embeddings of edges. The embedding enhancer exploits temporal structural correlation and reconstructs multi-head attention to achieve noise-resistant enhancement of basic hyper-embeddings. Extensive experiments show that our proposed HADF can realize resistance to structural noises and outperform state-of-the-art methods in identifying anomalous edges in dynamic noisy graphs.