A fast projected conjugate gradient algorithm for training support vector machines
Tong Wen, Alan S. Edelman, David Gorsich · Contemporary mathematics - American Mathematical Society · 2003
Support Vector Machines (SVMs) are of current interest in the solution of classification problems. However, serious challenges appear in the training problem when the training set is large. Training SVMs involves solving a linearly constrained quadratic programming problem. In this paper, we present a fast and easy-to-implement projected Conjugate Gradient algorithm for solving this quadratic programming problem.Compared with the exiting ones, this algorithm tries to be adaptive to each training problem and each computer's memory hierarchy. Although written in a high-level programming language, numerical experiments show that the performance of its MATLAB implementation is competitive with that of benchmark C/C++ codes such as SVMlight and SvmFu. The parallelism of this algorithm is also discussed in this paper.