Shapelet Selection for Efficient Time Series Classification by Dynamic Time Warping

Abstract

Shapelets are often used to solve time series classification problems and have a major drawback that they require high computational complexity in the extraction process. In order to solve this problem, many researchers have developed methods to obtain shapelets efficiently, but their classification performances are not good or they require hyperparameters. In this study, we propose a shapelet selection method using DTW(dynamic time warping). The proposed method searches for frequent patterns occurring in time series through the warping path of DTW and uses it as shapelets. To validate the proposed method, twenty-one benchmark datasets of time series are applied to our method and the existing methods, with which the classification accuracy and shapelet extraction time are compared. The proposed method shows no significant difference from the previous studies in computation time, while attains excellent performance in classification accuracy.