A General KPI Anomaly Detection Using Attention Models

To determine whether a Web service executes accurately, IT operation engineers must manually monitor multiple KPIs (Key Performance Indications). KPI anomaly detection is crucial to make sure undisrupted business. However, it is a burdensome task for IT operations engineers that analyze multivariate KPIs and detect the KPI exceptions. As a multiple time series, the KPIs of a service has two kinds of dependencies: temporal dependence and intermetric dependence. Existing forecasting-based models usually focus on temporal dependence and ignore intermetric dependencies between different KPI dimensions, which leads to low accuracy in the multivariate KPIs anomaly detection. Therefore, we propose a model, named GGIAnomaly, to consider both temporal dependence and inter-metric dependence in KPI anomaly detection. GGIAmomaly is composed of four parts: KPIs pre-process, intermetric dependence processing, temporal dependence processing and anomaly detection. Specially, GAT (Graph Attention Network) is used in GGIAnomaly for capturing the relationship between different KPI sequences. To better model various patterns of temporal dependence, GGIAnomaly integrates a recent attention model, named Informer, with GRU (Gated Recurrent Unit Network). The experiments on open-source datasets show that GGIAnomaly has better performance on both univariate and multivariate KPI anomaly detection compared to the existing methods.