Abstract: Support Vector Machine (SVM) is an effective classification method. Because different feature selection algorithm on the impact of classifier is not the same, combined with the characteristics of SVM , an feature selection algorithm based on the largest margin of SVM is presented. Taking advantage of the new feature selection algorithm to test feature selection and simulation on the Iris data set, the experiment results showed that the new feature selection algorithm can not only remove the noise data effectively, but also enhance the classifier's generalization ability.