Piecewise statistic approximation based similarity measure for time series

Abstract

In the research field of time series analysis, dynamic time warping distance (DTW) is a prevalent similarity measure with high precision. However, the computational complexity of DTW is high, which makes it difficult to be applied to the high dimensional time series. An effective solution is to compute DTW on the piecewise representation (PR-DTW), which employs the features of the subsequences of time series for similarity measure. However, the features that most existing piecewise representations focus on are too simple, which capture only one aspect of the fluctuation information of time series, and thus influence the precision of PR-DTW. In order to solve this problem, we propose a novel piecewise representation model, named piecewise statistic approximation (PSA), for supporting the PR-DTW measure. Rather than focusing on a single type of features, PSA extracts multiple statistical features to capture the synthetic fluctuation information for similarity measure. Besides, by taking the weighted Euclidean distance for the subsequence matching in the subroutine, PSA based DTW (PSADTW) can discriminate the expressivities of the multiple features. Comprehensive experiments over 45 real-world datasets empirically demonstrate that, PSA is well suited to support both precise and efficient PR-DTW measure.