Combining the Global and Partial Information for Distance-Based Time Series Classification and Clustering

Abstract

Many time series representation schemes for classification and clustering have been proposed. Most of the proposed representation focuses on the prominent series by considering the global information of the time series. The partial information of time series that indicates the local change of time series is often ignored. Recently, researches shown that the partial information is also important for time series mining. However, the combination of these two types of information has not been well studied in the literature. Moreover, most of the proposed time series representation requires predefined parameters. The classification and clustering results are considerably influenced by the parameter settings, and, users often have difficulty in determining the parameters.
We attack above two problems by exploiting the multi-scale property of wavelet decomposition. The main contributions of this work are: (1) extracting features combining the global information and partial information of time series (2) automatically choosing appropriate features, namely, features in an appropriate wavelet decomposition scale according to the concentration of wavelet coefficients within this scale. Experiments performed on several benchmark time series datasets justify the usefulness of the proposed approach.