The similarity measure is a key operation in the analysis and mining of time-series data. One of the most popular and effective measures is Dynamic Time Warping (DTW). Particularly, in the time-series classification (TSC) domain, DTW has been extensively studied over the past two decades. Consequently, several improved versions have been proposed in the literature. A critical observation is that most of these variants have never been evaluated together in the context of TSC. In our opinion, we believe that there is a need to compare DTW’s variants under a unified framework. Moreover, we also believe that such a study is of fundamental importance and could drive meaningful conclusions for both researchers and practitioners. Our objective is to provide a comprehensive comparison in which we show which variant is the most suitable for a particular problem. In this paper, we conduct an extensive evaluation to compare the classical DTW and its most popular variations for TSC. We evaluate these methods in terms of classification accuracy using a large variety of data-sets from the UCR time-series archive. The results show that no variant outperforms the others for all problems. Results also show that there is no statistically significant difference between virtually all variants.
Primary Language | English |
---|---|
Subjects | Software Engineering (Other) |
Journal Section | Articles |
Authors | |
Publication Date | June 5, 2021 |
Acceptance Date | December 16, 2020 |
Published in Issue | Year 2021 Volume: 4 Issue: 1 |
International Journal of Informatics and Applied Mathematics