Tri-Partition State Alphabet-Based Sequential Pattern for Multivariate Time Series

Recently, the advancement of cognitive computing and three-way decisions has enabled in-depth sequential pattern understanding through temporal association analysis. The main challenge is to obtain concise patterns that express richer semantics for multivariate time series (MTS) analysis. In this paper, we propose a tri-partition state alphabet-based sequential pattern (Tri-SASP) for MTSs. First, a tri-wildcard gap inserted between each pair of adjacent states enhances the flexibility of the method. Second, a given set of states is partitioned into positive (POS), negative (NEG) and boundary (BND) regions. The states in POS can only be used to construct a Tri-SASP, the states in NEG can only be matched by a tri-wildcard gap, and the states in BND can be used in both ways. Finally, horizontal and vertical algorithms are proposed to obtain frequent Tri-SASPs in a breadth-first manner. The experimental results on four real-world datasets show that (1) the discovered Tri-SASPs and temporal rules can enrich human cognition; (2) the two tri-partition strategies can bring us very meaningful and varied Tri-SASPs; and (3) the two algorithms are effective and scalable.