The performance of a combined distance between time series

Abstract

This paper presents the comparison of a proposed measure of dissimilarity between time series (COMB) with three baseline measures. COMB is a convex combination of Euclidean distance, a Pearson correlation based distance, a Periodogram based measure and a distance between estimated autocorrelation structures. The comparison resorts to 1-Nearest Neighbour classifier (1NN) since the effectiveness of the dissimilarity measures is directly reflected on the performance of 1NN. Data considered is available in the University of California Riverside (UCR) Time-Series Archive which includes data sets from a wide variety of application domains and have been used in similar studies. The COMB measure shows promising results: a good trade-off performance-computation time when compared to the alternative distances considered.