A robust pruning method for reduced weighted least squares support vector machine

In this paper, a robust pruning method for reduced weighted least squares support vector machine (RW-LSSVM) is proposed in order to improve the sparseness simultaneity concerning the robustness. By replacing the inequality constraints in SVM with equality constraints, LSSVM has a quite lower computation complexity meanwhile leading to a lack of robustness and sparseness. Several algorithms have been proposed to solve these problems but they usually focus on only one aspect. As always facing the situation that a well-performed sparse solution has an inferior generalization performance especially for datasets with high noises, the RW-LSSVM algorithm is considered as a proper way to robustly prune the support vectors to get a sparse solution with generalization considered as well. This algorithm set weights to the whole training set to ensure robustness and in each pruning step it uses the reduced LSSVM form to get a satisfactory sparseness by concerning the influence of all training samples on the sparse model where the model is composed only by a part of samples namely support vectors. To show the efficacy and feasibility of our proposed algorithm, some comparing experiments are carried out which are all favorable for our viewpoints.