Locally linear approximation for Kernel methods : the Railway Kernel

In this paper we present a new kernel, the Railway Kernel, that works properly for general (nonlinear) classification problems, with the interesting property that acts locally as a linear kernel. In this way, we avoid potential problems due to the use of a general purpose kernel, like the RBF kernel, as the high dimension of the induced feature space. As a consequence, following our methodology the number of support vectors is much lower and, therefore, the generalization capability of the proposed kernel is higher than the obtained using RBF kernels. Experimental work is shown to support the theoretical issues.