Accurate and efficient classification based on common principal components analysis for multivariate time series

Abstract

Multivariate time series are found everywhere and they are important data in the field of data mining, but their high dimensionality often hinders the quality of techniques employed for classifying multivariate time series. In this study, we propose an accurate and efficient classification method based on common principal components analysis for multivariate time series. First, multivariate time series are divided into several clusters according to the number of class labels, and the high dimensionality of multivariate time series can then be reduced by common principal components analysis, which gives the reduced principal component series sufficiently high variance. Second, each cluster is used to construct the corresponding reduced coordinate space formed by the eigenvectors of the common covariance matrix. Third, any multivariate time series without a class label can be projected onto these coordinate spaces and its label can be predicted based on the minimal variance of the reduced principal components series according to the different projections. Our experimental results demonstrated that the proposed method for the classification of multivariate time series is more accurate and efficient than existing methods. It is also flexible for multivariate time series with different lengths.