Multiscale Feature Extraction for Time Series Classification with Hybrid Feature Selection

Abstract

Time series classification has attracted increasing interests in recent years. Time series data are normally high dimensional data and it is well known that high dimensionality decreases the classification accuracy. Many feature extraction algorithms have been proposed to reduce the dimensionality of time series as a preprocessing stage. However, most of the proposed feature extraction algorithms don’t utilize the relevance between features and classes. Moreover, the optimized dimensionality often needs to be given by the users that is not convenient in real cases. In this paper, we propose a feature extraction algorithm for time series classification utilizing the relevant information of features and classes. The key idea of this algorithm is combining unsupervised multiscale feature extraction with supervised feature selection techniques. The optimized dimensionality could be detected directly from the algorithm. Experimental results on several benchmark time series datasets demonstrate the benefits of the proposed algorithm.