Characteristics based dissimilarity measure for time series

Abstract

It is necessary to invent dissimilarity measures which take into account the temporal nature of a time series. Such measures can be utilized for classification and clustering of time series. Great work has been conducted on this problem, but most measures use dimensionality reduction techniques. Such methods give accurate results for big data, but demonstrate a weakness now in short time series clustering. Many fields such as economics, demography, sociology, and others are presented by short time series. That is why a new method based on time series characteristics is introduced here. It is based on time series characteristics which are split into three groups: constant, dynamic and behavioural. A researcher can control the influence of the characteristics of each group as a result. Besides, we present a brief description of up-to-date dissimilarity measures from the R environment. The results of experiments on two synthetic data sets and comparison of our measure and other up-to-date methods are then presented. Refs 12. Figs 2. Table 1.