Summary
This work explores the capacity of Stacking to generate multivariate time series classifiers from classifiers of their univariate time series components. The Stacking scheme proposed uses k-nearest neighbors (K-NN) with dynamic time warping (DTW) as a dissimilarity measure for the level 0 learners. Support vector machines and Naïve Bayes are applied at level 1. The method has been tested on two data sets: Continuous plant diagnosis and Japanese vowels. Experimental results show that for these data sets the proposed Stacking configuration performs well when multivariate DTW fails to produces precise K-NN classifiers, increasing the accuracy achieved by K-NN as a stand alone method by the order of magnitude. This is an interesting issue because good univariate time series classifiers do not always perform satisfactory when adapted to the multivariate case. On the contrary, if the multivariate classifier is accurate, Stacking univariate classifiers may perform worse.