Recently, [1],we proposed a newdissimilarity measure between time series - COMB, a uniform convex combination of four (normalized) distance measures: Euclidean; Pearson correlation based; Periodogram based; and a distance between estimated autocorrelation structures. In this work, we propose a method to determine the weights of the convex combination of distances in COMB: it relies on the concordance of clusterings obtained by each individual distance measure and COMB derived clustering. A weighted COMB measure is thus obtained, WCOMB. We then test the clustering performance of WCOMB vs. COMB by conducting an experimental analysis on all the time series datasets of the UCR archive. We evaluate the concordance between the clusters obtained using K-Medoids and the original classes (using adjusted Rand index) as well as the cohesion-separation of the clusters (using the Silhouette index). In addition, we consider a clustering application - with data from the Portuguese Transmission System Operator, on time series of electricity consumption (2014 to 2019) - to compare the performance of both methods. Significant differences between the average Silhouette values of clusters obtained were found. The concordance with the original classes’ structure exhibits similar performance in both approaches. We conclude that, for unsupervised leaning, it can be worthwhile to invest on deriving specific weights for the distances integrating COMB.

--

clustering,distance measures,time series