Instance-Level Cost-Sensitive Hypergraph Learning With Quality-Aware Structure Judgement for Anomaly Detection

Anomaly detection aims to identify data instances that deviate from normal patterns or exhibit abnormal characteristics compared with typical samples. Due to the superiority in modeling high-order relationships, hypergraph has been widely adopted in anomaly detection tasks. However, in practical applications, different data instances may have varying influence on the detection process, and the effectiveness of hypergraph-based methods largely relies on the quality of the constructed hypergraph. To address these challenges, we propose an instance-level cost-sensitive hypergraph learning method with quality-aware structural judgement (ICSHL) for anomaly detection. Specifically, ICSHL integrates instance-level cost information into the hypergraph construction, allowing the model to capture the varying importance of samples and build a instance-level cost-sensitive hypergraph. Furthermore, ICSHL preserves the structural quality of the hypergraph by emphasizing structures with large margin separations while suppressing those with small margins. To evaluate the performance of the proposed method, we conduct extensive experiments on three widely used anomaly detection datasets. The experimental results and comparisons with state-of-the-art methods demonstrate the effectiveness and superiority of our approach.