Improved Iterative Sparseness for Least Squares Support Vector Machines

An improved iterative sparse algorithm is proposed to accelerate the execution of sparse least squares support vector machines(LS-SVM). Firstly, the technique of iterative approximation to the L0-norm is used to sparsify the LS-SVM for regression. However, each iteration requires solving a linear system with expensive computation compared to training a single LS-SVM. In this paper, improved conjugate gradient (ICG) method is given to reduce the computational cost, which is based on transforming the constrained primal problem in LS-SVM into an unconstrained minimization problem. Then the solution to the unconstrained minimization problem is obtained by using the CG method only once at each iteration. Finally, the result of numerical experiment shows that the proposed method get sparse LS-SVM model with significant reduction in computational cost