A Scale-Space Theory and Bag-of-Features Based Time Series Classification Method

The aim of this study is to develop a time series classification method based on scale-space theory. Our study has been conducted in three steps: In the first step, scale-space extrema of time series found through using SiZer (Significant ZERo crossings of the derivatives) method and local features set constructed around the determined extreme points, based on interval-widths list entered by the user. In the second step, the values of descriptors have been computed around the prescribed scale-space extrema. In this study we have used mean, standard deviation and the slope of fitted regression line as descriptors for each interval and with the aid of these values bag-of-features has been constructed. In the third and the last step, after the obtained bag-of-features set clustered, the classification procedure has been completed by using random forest method. Error rates of the proposed method have been compared with the error rates of some widely-used methods by using UCR time series database and it is concluded that the obtained results are better by a majority. It is planning to take forward our study by amendment of the method for finding scale-space extrema and including other descriptors.