Abstract: complexity of SVM is highly dependent on size of data set. Granular support vector machine (GSVM) produces the reduced training set by selecting representative points of granules in order to improve the learning efficiency of support vector machine (SVM). However, the selection of individual representative points may likely lose part of the important classification information, which will lead to poor model generalization performance. To solve this problem, this paper presents a distribution based GSVM (DGSVM) to accelerate the training process. It selects representative points according to the distribution of positive and negative samples in the granules, which will be divided into the reduced training sample set and modified set by the measure of mixing degree of granules. Then the initial classifier can be obtained by the reduced training sample set, and the modified set is used to further optimize the classifier. Naturally, DGSVM obtains initial classifier with those difficult classification samples, and then modify classifier gradually. The experiment results on UCI benchmark datasets demonstrate that DGSVM model can improve the generalization performance of greatly with high learning efficiency synchronously