Adaptive sequential three-way decisions for dynamic time warping

Abstract

Dynamic time warping (DTW) algorithm is widely used in diversified applications due to its excellent anti-deformation and anti-interference in measuring time-series based similarity. However, the high time complexity of DTW restrains the applicability of real-time case. The existing DTW acceleration studies suffer from a loss of accuracy. How to accelerate computation while maintaining satisfying computational accuracy remains challenging. Motivated by sequential three-way decisions, this paper develops a novel model with adaptive sequential three-way decisions for dynamic time warping (AS3-DTW). Firstly, we systematically summarize distance differences under the context of adjacent tripartite search spaces for DTW, and propose five patterns of granularity adjustments of the search spaces. Furthermore, we present the corresponding calculation method of DTW adjacent tripartite search spaces distances difference. Finally, we construct a novel dynamism on adaptively adjusting time warping by combining sequence-based multi-granularity with sequential three-way decisions. Experimental results show that AS3-DTW effectively achieves promising trade-off between computational speed and accuracy on multiple UCR datasets when compared with the state-of-the-art algorithms.