Support Vector Machines have been promising tools for data mining during these years because of their good performance. However, a main weakness of SVMs is lack of comprehensibility: people can not understand what the “optimal hyperplane” means and are unconfident about the prediction especially when they are not the domain experts. In this paper we introduce a new method to extract knowledge with a thought inspired by the decision tree algorithm and give a formula to find the optimal attributes for rule extraction. The experimental results will show the efficiency of our algorithm.