A sparsity driven kernel machine based on minimizing a generalization error bound

A new sparsity driven kernel classifier is presented based on the minimization of a recently derived data-dependent generalization error bound. The objective function consists of the usual hinge loss function penalizing training errors and a concave penalty function of the expansion coefficients. The problem of minimizing the non-convex bound is addressed by a successive linearization approach, whereby the problem is transformed into a sequence of linear programs. The algorithm produced comparable error rates to the standard support vector machine but significantly reduced the number of support vectors and the concomitant classification time.