Frequent State Transition Patterns of Multivariate Time Series

Sequence pattern discovery is a key issue in multivariate time series analysis. Popular approaches first obtain the pattern of each single-variate time series and then obtain cross-variate associations. In this paper, we consider different variables at the same time during pattern construction, and propose a new type of pattern called State Transition pAttern with Periodic wildcard gaps (STAP). Compared to previous types, STAP reveals stronger cross associations among different variables and provides better interpretability for decision makers. We design an approach with two stages, namely frequent state discovery and pattern synthesis, to obtain frequent STAPs. We propose two pre-pruning and an Apriori-pruning techniques to speed up pattern discovery. We also propose two post-pruning techniques to simplify the output and a visualization way to support expert decision. Experimental results on four real-world datasets demonstrate 1) STAP captures the cross and temporal associations; 2) the five pruning and pattern synthesis techniques are quite effective; and 3) visualization technique greatly increases the readability of STAP.