Utilizing Multivariate Time Series for Semantic Segmentation

The rise of connected devices and new communication technologies has resulted in an enormous wealth of multivariate time series data. The identification and extraction of meaningful segments by means of data mining algorithms has many applications. The problem of multivariate time series segmentation has been studied extensively with statistical methods that rely on the statistical properties of the time series for segmentation. We introduce a novel method, which exploits domain-specific information from the multivariate time series for segmentation. As a proof-of-principle, we demonstrate the feasibility of our method. Results show that after segmentation, the running time of anomaly detection algorithms reduces significantly, while preserving the effectiveness of anomaly detection.