Efficient Dynamic Time Warping for Time Series Classification

Abstract

Background/Objective: Dynamic Time Warping (DTW), a similarity measure works in O(n2) complexity. Cause of this it will be used for small datasets only. Methods/Statistical Analysis: In this work, we introduced Efficient DTW (EDTW), which works in linear time. It uses two level approaches. In the first level data reduction is performed, and in the second level warping distance and path are calculated. Findings: While calculating the values of distance matrix, values along the warping path only considered and calculated. For time series of length n, maximum n values of distance matrix are calculated. So it works in linear time. Improvements/Applications: We applied this distance measure to UCR Time Series archive and calculated error rate of 1NN classification. Most of the cases it is matching, some cases it is better, and some other cases error rate is high.