Hybrid Anomaly Detection via Multihead Dynamic Graph Attention Networks for Multivariate Time Series

In the real world, a large number of multivariate time series data are generated by Internet Things systems, which composed many connected sensing devices. Therefore, it is impractical to consider only single univariate for decision-making. High-dimensional decrease performance traditional anomaly detection methods. Moreover, previously developed methods capture temporal correlations instead spatial correlations. necessary learn and between different timestamps. this paper, achieve improved series, we propose novel architecture based on graph attention network (GAT) with multihead dynamic (MDA). This framework simultaneously learns dependencies sensors in both dimensions. To tackle overfitting problem autoencoder (AE)-based methods, hybrid approach that combines generative adversarial (GAN) as reconstruction model multilayer perceptron (MLP) prediction-based detect anomalies together. The proposed paper called HAD-multihead GAT (MDGAT). Extensive experiments public benchmarks demonstrate superior HAD-MDGAT over state-of-the-art