Effects of Time Normalization on the Accuracy of Dynamic Time Warping

This paper revisits Dynamic Time Warping, a method for assessing the dissimilarity of time series. In particular, this paper provides theoretical and experimental evidence showing that uncritical normalizing the length of the time series to be compared has a detrimental effect on the recognition accuracy in application domains such as on-line signature recognition, where the length of compared time series matters for their classification as match or non-match.