Abstract: Two improvements are introduced into vicinal-risk-minimization based support vector algorithm.Since the misclassified samples must be support vectors,a scheme for pruning hard-to-learn samples from the training set based on support vectors is presented.The parameter's determination of Gaussian vicinal function is proposed to be modified,based on the maximum likelihood criterion.Preliminary experimental results show that the pruning scheme and improvement of the parameter's determination of vicinal function much improved Vicinal SV algorithm's generality,and can outperform SVM by about 0.5% in test accuracy.