Abstract: Based on One-Class support vector machine, Hyper-sphere support vector machine is an efficient tool to solve problems of multi-class classification. With theoretical deduction and experiments, we found that hyper-sphere support vector machine is over dependent on each training sample, even if the training sample belongs to noise or outlier. So a preprocessing method before training is proposed in this paper to remove noise and outlier by similarity computation from the original training data set. In the training process, a weight factor is added for each sample to differentiate its contribution to the resulting classifier, ensuring quick convergence of the algorithm based on SMO solving. In the testing process, an appropriate classification rule is selected by the test sample position. Experimental results compared to other hyper-sphere support vector machines, new proposed algorithm reduces influences of noises or outliers for the resulting classifier and improves classification accuracy.