RBF networks with dynamic barycenter averaging kernel for time series classification

Abstract

Radial basis function (RBF) network has received many considerable realistic applications, due to its simple topological structure and strong capacity on function approximation. However, the core of RBF network is its static kernel function, which is based on the Euclidean distance and cannot be directly used for time series classification (TSC). In this paper, a new temporal kernel called Dynamic Barycenter Averaging Kernel (DBAK) is introduced into RBF network. Specifically, we first determine the DBAK’s centers by combining k-means clustering with a faster DTW-based averaging algorithm called DTW Barycenter Averaging (DBA). And then, to allow the stable gradient-training process in the whole network, a normalization term is added to the kernel formulation. By integrating the warping path information between the input time series and the centers of kernel, DBAK based RBF network (DBAK-RBF) can efficiently work for TSC tasks. Experimental results demonstrate that DBAK-RBF can achieve the better performance than previous models on benchmark time series datasets.