Dynamic Time Warping of Segmented Time Series

Abstract

Providing the most suitable time series representation has always been a crucial factor in time series data mining. The selected approximation does not only determine the tightness of the representation but also the (dis)similarity measure to be used. Piecewise Linear Representation (PLA) is one of the most popular methods when tight representation of the original time series is required; however, there is only one Dynamic Time Warping (DTW) based dissimilarity measure which uses the PLA representations directly. In this paper, a new dissimilarity measure is presented which takes not only the mean of a segment into account but combines it with our recently introduced slope-based approach, which was derived from Principal Component Analysis (PCA).