Fast and accurate template averaging for time series classification

Time series data are evidently ubiquitous, as we could see them in all kinds of domains and applications. As a result, various data mining tasks are often performed to discover useful knowledge, including commonly performed tasks like time series classification and clustering. Dynamic Time Warping (DTW) is accepted as one of the best available similarity measures, which has been used for distance calculation in both classification and clustering algorithms. However, its known drawback is its exceedingly high computational cost. Recently, data condensation method through template averaging is applied; each class of data can be represented by one template which could greatly speed up the classification with DTW especially in large datasets, with the trade off in lower classification accuracies. Subsequently, various attempts have been made to increase the number of representative templates to boost up the accuracies while keeping the computation complexity not too high. However, those algorithms still suffer from many predefined and hard-to-set parameters, while some require high computation time for high accuracy results. Therefore, in this work, we propose an accurate yet simple template averaging method that is parameter free and has much less computation time. The experiment results on 20 UCR time series benchmark datasets demonstrate that our proposed method can achieve a few orders of magnitude speedup while maintaining high classification accuracies.