Abstract: Although the algorithm of linear support vector machine (LSVM) is simple,efficient in training and testing speeds,it can not be applied for nonlinear datasets．For overcoming its drawback,the original training data was splited into several subsets and their LSVMs were respectively constructed．Then,we fit a nonlinear decision function for solving linear inseparation through the combination of the nonlinear radical basis functions (RBFs)．Based on this motivation,we developed a new model,called fast model of ensembling LSVMs (FMELSVM),which is suitable for the classification of large datasets．This model improves the nonlinear capabilities of LSVMs using RBF．Meanwhile,the ensembling effects are enhanced by introducing an optimized weight vector．Experimental results on UCI demonstrate that FMELSVM obtains competitive effectiveness for large datasets．

Key words: Classification,Linear SVM,Radical basis function,Gradient descent method