DGNN: Dynamic Graph Neural Networks for Anomaly Detection in Multivariate Time Series
Bowen Chen · Proceedings/Proceedings of the ... International Conference on Software Engineering and Knowledge Engineering · 2023
In recent years, there has been significant progress in the importance of anomaly detection for multivariate time series in industrial applications.However, there are still limitations.Although deep learning methods have improved anomaly detection in high-dimensional multivariate time series, they are computationally expensive and do not explicitly learn the relational structure between sequences.In this paper, we propose an unsupervised anomaly detection algorithm called Dynamic Graph Neural Networks (DGNN).Firstly, we propose a datadriven method of generating "subgraphs" to interpret interior correlations between sequences, instead of using the traditional method of a fully connected graph.Secondly, we introduce a novel Graph Attention Networks based on correlation to fuse neighbor sequence features.Experimental results on five public datasets demonstrate that our method consistently achieves state-of-theart performance compared to other baseline methods, while reducing the edges of the graph by nearly 70%.Index Terms-time series, anomaly detection, graph attention network, correlation coefficient• We propose DGNN, a dynamic graph neural network approach that enables faster and more accurate learning of