The classification result of classical support vector machine algorithm in the case of unbalanced data set is not satisfactory, so a novel under-sampling algorithm based on sample properties is presented. According to sample information in the kernel space, a certain percentage of majority instances located near the classification interface are selected. Then according to the sample’s density, the representive majortiy samples in the selected samples are selected, which can not only reduce the number of majority instances, but also make the SVM classification interface bias toward the majority instances. In the experiments, the proposed approach is compared with other data-preprocess methods for unbalanced dataset classification, the experimental results demonstrate that the proposed method can improve classification performance of SVM in the minority class data, the overall classification performance and robustness.