A Non-parametric Wavelet Feature Extractor for Time Series Classification

Many representation schemes for time series have been proposed and most of them require predefined parameters. In case of classification, the accuracy is considerably influenced by these predefined parameters. Also, the users usually have difficulty in determining the parameters. The aim of this paper is to develop a representation method for time series that can automatically select the parameters for the classification task. To this end, we exploit the multi-scale property of wavelet decomposition that allows us to automatically extract features and achieve high classification accuracy. Two main contributions of this work are: (1) selecting features of a representation that helps to prevent time series shifts, and (2) choosing appropriate features, namely, features in an appropriate wavelet decomposition scale according to the concentration of wavelet coefficients within this scale.