Contrast enhanced dynamic time warping distance for time series shape averaging classification

Abstract

Dynamic Time Warping (DTW) distance has been a focal point of time series analysis in the past few years. Time series data mining such as classification and clustering normally requires shape averaging to create a template to represent a class. The DTW distance typically provides the similarity measure and the alignment between two time series. The DTW similarity measure can be used as a distance measure in various classification algorithms while the alignment provides the basis for time series shape averaging. Although the original DTW distance provide better similarity measure than the Euclidean distance, some accuracy improvements remain a challenge. Recently, many novel DTW varieties have been reported with greater accuracy and higher performance such as the Resampled DTW, Hybrid DTW, etc. To improve the accuracy further, we speculate that, in the DTW-based distance measurements, some negligible data points may have non-trivial contribution to the measured distance. Hence, we propose a novel Contrast Enhanced Dynamic Time Warping (CEDTW) that reduces the effect from those negligible data points and improves the accuracy in similarity measure. The accuracy of the new method is validated against the classic NLAAF, PSA and the latest RSA based shape average on classification problems. Copyright © 2009 ACM.