A New Perspective on Convex Relaxations of Sparse SVM

This paper proposes a convex relaxation of a sparse support vector machine (SVM) based on the perspective relaxation of mixed-integer nonlinear programs. We seek to minimize the zero-norm of the hyperplane normal vector with a standard SVM hinge-loss penalty and extend our approach to a zeroone loss penalty. The relaxation that we propose is a second-order cone formulation that can be efficiently solved by standard conic optimization solvers. We compare the optimization properties and classification performance of the second-order cone formulation with previous sparse SVM formulations suggested in the literature.