Multivariate time series clustering based on common principal component analysis

Time series clustering is often applied to pattern recognition and also as the basis of the tasks in the field of time series data mining including dimensionality reduction, feature extraction, classification and visualization. Due to the high dimensionality of multivariate time series and most of the previous work concentrating on univariate time series clustering, a novel method which is based on common principal component analysis, is proposed to achieve multivariate time series clustering more fast and accurately. It is inspired by the traditional clustering method K-Means and can construct a common projection axes as prototype of each cluster. Moreover, the reconstruction error of each multivariate time series projected on the corresponding common projection axes are used to reassign the member of the cluster. The detailed algorithm of the proposed method Mc2PCA is given and the time complexity is analyzed, which shows that the proposed method is very fast and its time complexity is linear to the number of multivariate time series objects. Unlike the traditional methods, the proposed method considers the relationship among variables and the distribution of the original data values of multivariate time series. The experimental results in the various datasets demonstrate that Mc2PCA is superior to the traditional methods for multivariate time series clustering. (C) 2019 Elsevier B.V. All rights reserved.