Learning optimal warping window size of DTW for time series classification

Abstract

The dynamic time warping (DTW) is a classic similarity measure which can handle the time warping issue in similarity computation of time series. And the DTW with constrained warping window is the most common and practical form of DTW. In this paper, the traditional learning method for optimal warping window of DTW is systematically analyzed. Then the time distance to measure the time deviation between two time series is introduced. Finally a new learning method for optimal warping window size based on DTW and time distance is proposed which can improve DTW classification accuracy with little additional computation. Experimental data show that the optimal DTW with best warping window get better classification accuracy when the new learning method is employed. Additionally, the classification accuracy is better than that of ERP and LCSS, and is close to that of TWED.