Sparse Least Squares Low Rank Kernel Machines

Abstract

A general framework of low rank least squares support vector machine (LR-LSSVM) is introduced in this paper. The special structure of controlled model size of the low rank kernel machine brings in remarkable sparsity and hence gigantic breakthrough in computational efficiency. In the meantime, a two-step optimization algorithm with three regimes for gradient descent is proposed. For demonstration purpose, experiments are carried out using a novel robust radial basis function (RRBF), the performances of which mostly dominate.