Parameter-free motif discovery for time series data

Abstract

Time series motif discovery is an increasingly popular research area in time series mining whose main objective is to search for interesting patterns or motifs. A motif is a pair of time series subsequences, or two subsequences whose shapes are very similar to each other. Typical motif discovery algorithm requires a predefined motif length as its parameter. Discovering motif with arbitrary lengths introduces another problem, where selecting a suitable length for the motif is non-trivial since domain knowledge is often required. Thus, this work proposes a parameter-free motif discovery algorithm called k-Best Motif Discovery (kBMD) which requires no parameter as input, and as a result returns a set of all “Best Motif” that are ranked by our proposed scoring function which is based on similarity of motif locations and similarity of motif shapes.