Abstract
The support vector classification-regression machine for K-class classification (K-SVCR) is a novel multi-class classification method based on 1-versus-1-versus-rest structure. In this paper, we propose a least squares version of K-SVCR named as LSK-SVCR.
Similarly as the K-SVCR algorithm, this method assess all the training data into a 1-versus-1-versus-rest structure, so that the algorithm generates ternary output -1,0,+1. In LSK-SVCR, the solution of the primal problem is computed by solving only one system of linear equations instead of solving the dual problem, which is a convex quadratic programming problem in K-SVCR.
Experimental results on several benchmark data set show that the LSK-SVCR has better performance in the aspects of predictive accuracy and learning speed.