Local anomaly detection via subgraph embedding on multivariate data streams

This study addresses the crucial challenge of local anomaly detection in multivariate data streams, which is essential for applications such as fault detection and damage localization in industrial contexts. Traditional approaches often neglect two critical aspects: the significance of relational features in assessing similarity and the necessity of pinpointing specific data streams that exhibit abnormal patterns. To address these gaps, we introduce a novel subgraph embedding-based method. Our approach constructs dynamic graphs to model relational features and identifies anomalies by detecting subgraphs within these graphs. The core innovation lies in an entropy-based transformer integrated with an autoencoder, which maps subgraphs into a low-dimensional space to effectively discriminate abnormal subgraphs. Experiments on EEG and industrial datasets show that the proposed model achieves average F1-score improvements of 0.07 and 0.105 over the compared baselines, respectively. The results indicate that entropy-based subgraph embedding can improve local anomaly detection performance, although the model does not achieve the best value for every metric on every dataset.