A PCA-based similarity measure for multivariate time series

Abstract

Multivariate time series (MTS) datasets are common in various multimedia, medical and financial applications. We propose a similarity measure for MTS datasets, <i>Eros</i> <i>E</i>xtended F<i>ro</i>beniu<i>s</i> norm), which is based on Principal Component Analysis (PCA). <i>Eros</i> applies PCA to MTS datasets represented as matrices to generate principal components and associated eigenvalues. These principal components and eigenvalues are then used to compare the similarity between MTS matrices. Though <i>Eros</i> in itself does not satisfy the triangle inequality, without which existing multidimensional indexing structures may not be utilized, the lower and upper bounds to satisfy the triangle inequality are obtained. In order to show the validity of <i>Eros</i> for similarity search on MTS datasets, we performed several experiments on three datasets (2 real-world and 1 synthetic). The results show the superiority of our approaches as compared to the traditional similarity measures for MTS datasets, such as Euclidean Distance (ED), Dynamic Time Warping (DTW), Weighted Sum SVD (WSSVD) and PCA similarity factor (S<sc>PCA</sc>) in precision/recall.