A method for measuring similarity of time series based on series decomposition and dynamic time warping

Dynamic time warping (DTW) is one of the most important similarity measurement methods for time series analysis. In view of the high complexity and pathological alignment of DTW, a lot of variants of DTW have been proposed. However, the existing methods calculate the similarity between the original time series through dynamic programming directly, and ignore the characteristic that different components in the time series often have different degrees of importance. This paper proposes a time series similarity measurement method based on series decomposition and fast DTW, which combines time series decomposition method and DTW method. Series decomposition is an important means of time series analysis which can decompose time series into trend, seasonality, and remainder components. In this paper, after using the Seasonal-Trend decomposition using Loess (STL) method to decompose the time series, the similarity between the trend components and the similarity between seasonal components are respectively measured. The impact of the more important component is amplified, and then the comprehensive similarity measurement result will be obtained. Experimental results on 20 UCR time series datasets show that, compared with the existing fast DTW and constrained DTW and their variants, the method proposed in this paper achieves a higher classification accuracy. Simultaneously, combining the advantage of low complexity of fast DTW, the computational complexity of proposed method is still first-order linearly related to the length of time series.

Automated summary

Dynamic time warping (DTW) is one of the most important similarity measurement methods for time series analysis. This paper proposes a time series similarity measurement method based on series decomposition and fast DTW, which calculates the similarity between different components in the time series and amplifies the impact of the more important component to obtain a comprehensive similarity measurement result.