Multi-class support vector machine based on the minimization of class variance

摘要

Since the existing methods can not balance the sufficient use of information and the scale of the optimization problem, a new method for multi class classification problem is proposed, which is called multi-class support vector machine based on the minimization of class variance (MCVMSVM for short). MCVMSVM adopts the idea of semi-supervised learning and transfers the K-class problem to K(K − 1)/2 binary classification problems. For each binary classification problem, a new SVM with a mixed regularization term which considers the margin and the distribution of examples is proposed. MCVMSVM can utilize the information of all examples without increasing the scale of the optimization problem. The performance of MCVMSVM on UCI and NDC datasets is the best compared with other methods, that means MCVMSVM is more effective.