Averaging Methods using Dynamic Time Warping for Time Series Classification

Abstract

Averaging is an important step in time series classification or clustering, to create representative sequences for each category of data. A global averaging method for Dynamic Time Warping (DTW) based time series analysis is DTW Barycenter Averaging (DBA). In this paper, we propose a recursive tree based implementation of DBA, for faster computation of an average sequence, using the divide-and-conquer strategy. We also propose to automate the termination of DBA using a data-driven approach. The performance of the proposed methods is evaluated using accuracy, precision and recall as performance metrics, in a simple DTW-distance based classification method on ten standard time series datasets. Experimental results demonstrate that the proposed approaches are significantly faster than DBA, while achieving similar performance.