A novel sparse least-squares support vector machine

Classical Least-Squares Support Vector Machines (LS-SVM) severely suffer from non-sparseness problem. Previous methods address this issue by simplifying the decision rule post training, which risks a loss in generalization ability and impose extra computation cost. The paper proposed to apply a novel function approximation technique for the training of a binary Least Squares Support Vector Machine (LS-SVM). The novel training algorithm can detect the linear dependencies between vectors of the input Gram matrix, which eases the non-sparseness problem of the conventional LS-SVM. Experiments on two-spiral datasest illustrate that, the proposed LS-SVM can effectively produce an optimal hyperplane which is sparse in training examples.