Learning with multiple Gaussian distance kernels for time series classification

Various distance measures have been proposed for time series classification, and several of them have been used to construct Gaussian distance kernels for support vector machine (SVM) - based classification. Considering that different Gaussian distance kernels may carry complementary information for classification, in this paper, we propose a multiple kernel learning (MKL) method to integrate multiple Gaussian distance kernels to further improve time series classification accuracy. We first adopt the classical Gaussian RBF (GRBF) kernel and the recently developed Gaussian elastic metric distance kernel (i.e. GERP kernel and GTWED kernel), and then use an efficient MKL, SimpleMKL, to learn the kernel classifier. Our experimental results on 12 UCR time series data sets show that the proposed method is superior to SVM with individual Gaussian distance kernel.