Multivariate Time Series Anomaly Detection Using Directed Hypergraph Neural Networks

Multivariate time series anomaly detection is a challenging problem because there can be a number of complex relationships between variables in multivariate time series. Although graph neural networks have been shown to be effective in capturing variable-variable relationships (i.e., relationships between two variables), they are hard to capture variable-group relationships (i.e., relationships between variables and groups of variables). To overcome this limitation, we propose a novel method called DHG-AD for multivariate time series anomaly detection. DHG-AD employs directed hypergraphs to model variable-group relationships within multivariate time series. For each time window, DHG-AD constructs two different directed hypergraphs to represent relationships between variables and groups of positively and negatively correlated variables, enabling the model to capture both types of relationships effectively. The directed hypergraph neural networks learn node representations from these hypergraphs, allowing comprehensive multivariate interaction modeling for anomaly detection. We show through experiments using various evaluation metrics that our proposed method achieves the best scores among the compared methods on two real-world datasets.