Skip to Main Content
In this paper, in order to reduce the support vectors on a large scale data set, we train support vector machines which utilize the hyper-spheres as the training samples. By representing adjacent samples of the same class as hyper-spheres, the boundary location can be controlled both by the center and radius of the hyper-spheres. We demonstrate that the optimization problem in this condition can be solved easily only by revising initial conditions of sequential minimal optimization (SMO) algorithm. Compared with previous algorithms on several data sets, the proposed algorithm is quite competitive in both the computational efficiency and the classification accuracy.