Series2Graph++: Distributed Detection of Correlation Anomalies in Multivariate Time Series

Multivariate time series are a form of real-valued sequence data that simultaneously record different time-dependent variables. They originate mostly from multi-sensor setups and serve a variety of important analytical purposes, including the detection of normal and abnormal behavior. To efficiently detect both single-channel and correlation anomalies in time series of real-world size, we propose Series2Graph++ (S2G++), an unsupervised, distributable anomaly detection algorithm for wide and long multivariate time series. S2G++ extends the univariate S2G algorithm and its distributed variant DADS by adding support for multidimensional time series and, hence, correlation anomalies. For this, we translate S2G’s graph-based anomaly detection approach into multidimensional spaces. We additionally propose a root cause feature that serves to relate the detected anomalies to the anomalous channel(s). Our experiments demonstrate that S2G++ is significantly faster than related algorithms and still competes for the best quality results.