Common Dissimilarity Measures are Inappropriate for Time Series Clustering

Clustering algorithms have been actively used to identify similar time series, providing a better understanding of data. However, common clustering dissimilarity measures disregard time series correlations, yielding poor results. In this paper, we introduce a dissimilarity measure based on series partial autocorrelations. Experiments compare hierarchical clustering algorithms using the common dissimilarity measures, such as Euclidean Distance and Dynamic Time Warping, to cluster time series following Box-Jenkins Auto-Regressive models. Results show that our dissimilarity measure produces better results for both synthetic and real data sets in terms of the Adjusted Rand Index and Normalized Hubert Γ statistic. Our findings confirm that the choice of dissimilarity measure is crucial for improving time series clustering quality.