ABSTRACT
A fundamental problem in clustering and classification analysis is the choice of a relevant metric. Many feature-based methods have been developed to address the problem of clustering raw time series data. There are advantages associated with the feature-based clustering approach. In particular, dimensional reduction will in general be attained and this is likely to lead to saving on computational time. In addition, the procedures based on feature extraction can be commonly applied to time series of different lengths. Features extracted from stationary time series are not necessarily going to be the same as those extracted from non-stationary time series. Methods based on features extracted in the time domain, the frequency domain and from wavelet decomposition of the time series are presented. Frequency domain features Spectral ordinates Spectral analysis provides useful information about the time series behaviour in terms of cyclic patterns and periodicity.
