GPU-based time series classification through ensembles of non-elastic and elastic distances

Abstract

Improving effectiveness and efficiency of time series classification is a very important task. This paper presents a new ensemble approach for time series classification in which each 1-NN (one-nearest-neighbor) classifier coupled with a non-elastic or elastic distance measure. We also design a GPU-based parallel implementation for the proposed ensemble method in order to improve its efficiency.  Experimental results over a  collection of benchmark datasets showed that our proposed method remarkably outperforms the 1NN with Dynamic Time Warping measure which has been considered in literature as “difficult to beat” and brings out the same classification accuracy as the ensemble method which uses eight elastic distance measures (the Elastic Ensemble [14]). Besides, we compare the efficiency of GPU-based implementation to that of sequential implementation for the proposed method. Results in the latter experiment showed that in average the GPU version can run faster that the sequential version about 48 times.