Upper bounds for time and accuracy improvement of dynamic time warping approximation

Abstract

Dynamic time warping (DTW) consists at finding the best alignment between two time series. It was introduced into pattern recognition and data mining, including many tasks for time series such as clustering and classification. DTW has a quadratic time complexity. Several methods have been proposed to speed up its computation. In this paper, we propose a new variant of DTW called dynamic warping window (DWW). It gives a good approximation of DTW in a competitive CPU time. The accuracy of DWW was evaluated to prove its efficiency. Then the KNN classification was applied for several distance measures (dynamic time warping, derivative dynamic time warping, fast dynamic time warping and DWW). Results show that DWW gives a good compromise between computational speed and accuracy of KNN classification.