Research on dynamic time warping multivariate time series similarity matching based on shape feature and inclination angle

Danyang Cao,Jie Liu

doi:10.1186/s13677-016-0062-z

Abstract

Different sets of research mainly focus on one variable time series now, while researches involving multivariate time series have been insufficient. In this paper, combined linear segments and fitting error for multivariate time series, we present a new method to reduce the time complexity of DTW distance metric algorithm. Based on the shape feature and the tilt angle, we propose a new approach for similarity matching of DTW multivariate time series. Experimental results demonstrate that this method is helpful for ensuring accuracy and for reducing the time complexity of similarity matching.

Highlights

Time series are widely used in economics, management, computers, mathematics, electronics and many other interdisciplinary researches [1]
PAA_ERR: the segmentation method to multivariate time series When we use the PAA_ERR method on the multivariate time series, we take all the variables as a whole into consideration, that is, we use Piecewise Aggregate Approximation (PAA) to divide all variables into segments, if the value of overall fitting error etotal of a segment is greater than the threshold value on all the variables, reprocess the segment, until the etotal is less than threshold value. The steps of this algorithm are as follows: Assume that time series MTS processed by PAA_ERR can be divided into s segments, the linear representation of these multivariate time series after processing can be denoted as L(MTS) = {L(xi1, xi2), L(xi2, xi3), ..., L(xik- 1,xik),i∈[1,m],k∈[1,s]}, where i represents a variable of the segmentation, where xik represents the record value of the time series, L(xik- 1,xik) represents a straight line connecting two points
In the diagram, when the slope angle weight value was 0, the weight of Trend Distance (TD) method’s time span and SA_DTW’s shape mode value were 1 respectively, the accuracy of TD method decreased to 20 %, while SA_DTW still had high accuracy. It indicated that the shape weight we proposed in this paper was more suitable for the representation of multivariate time series model

Summary

Introduction

Time series are widely used in economics, management, computers, mathematics, electronics and many other interdisciplinary researches [1]. Multivariate time series model representation From the TD method of Lee et al, we can see that when the weight values corresponding to the appropriate tilt angle and the time maintaining length were changed, the accuracy rate would be changed significantly.

Results

Conclusion