A comparative study of similarity measures for time series classification

Abstract

Time series data are found everywhere in the real world and their analysis is needed in many practical situations. Multivariate time series data poses problem for analysis due to its dynamic nature and traditional machine learning algorithms for static data become unsuitable for direct application. A measure to assess the similarity of two time series is essential in time series processing and a lot of measures have been developed. In this work a comparative study of some of the most popular similarity measures has been done with 43 benchmark data set from UCR time series repository. It has been found that, on the average over the different data sets, DTW performs better in terms of classification accuracy but it has high computational cost. A simple processing technique for reducing the data set to lower computational cost without much degradation in classification accuracy is proposed and studied. A new similarity measure is also proposed and its efficiency is examined compared to other measures.