Robust classification of noisy data using second order cone programming approach

Assuming an ellipsoidal model of uncertainty a robust formulation for classifying noisy data is presented. The formulation is a convex optimization problem, in particular it is a instance of second order cone programming problem. The formulation is derived from a worst case consideration and the robustness properties hold for a large class of distributions. The equivalence of ellipsoidal uncertainty and Gaussian noise models is also discussed. The generalized optimal hyperplane is recovered as a special case of the robust formulation. Experiments on real world datasets illustrates the efficacy of the formulation.