Flexible Shapelets Discovery for Time Series Classification

Time series classification is important due to its pervasive applications, especially for the emerging Smart City applications that are driven by intelligent sensors. Shapelets are sub-sequences of time series that have highly predictive abilities, and time series represented by shapelets can better reveal the patterns thus have better classification accuracy. Finding shapelets is challenging as its computational in-feasibility, most existing methods only finds shapelets with a same length or a few fixed length shapelets because the searching space of shapelets with arbitrary length is too large. In this paper, we improve the time series classification accuracy by discovering shapelets with arbitrary lengths. We borrow the idea of Apriori algorithm in association rule learning, that is, the superset shapelet candidates of a poor predictive shapelet candidate also have poor predictive abilities. Therefore, we propose a Flexible Shapelets Discovery (FSD) algorithm to discover shapelets with varying lengths. In FSD, shapelet candidates having the lower bound of length are discovered, and then we extend them into arbitrary lengths shapelets as long as their predictive abilities increases. Experiments conducted on 6 UCR time series datasets demonstrate that the arbitrary length shapelets discovered by FSD achieves better classification accuracy than those using fixed length shapelets.