Abstract
Since the training of support vector machines needs to solve the dual problem with the number of variables equal to the number of training data, the training becomes slow when the number of training data is large. To speed up training the Sequential Minimal Optimization (SMO) technique has been proposed, in which two data are optimized simultaneously. In this paper, we propose to extend SMO so that more than two data are optimized simultaneously. Namely, we select a working set including variables, solve the equality constraint for one variable included in the working set, and substitute it into the objective function. Then we solve the subproblem related to the working set by calculating the inverse of the Hessian matrix. We evaluate our method for the five benchmark data sets and show the speed-up of training over SMO.
