For the multi-class classification with Support Vector Machines (SVMs), a decision tree architecture has been proposed for computational efficiency. But by SVM decision tree, the generalization ability depends on the tree structure. In this paper, to improve the generalization ability of SVM decision tree, a novel separability measure is defined based on the distribution of the training samples in the kernel space, and an improved SVM decision tree is provided. The theoretical analysis and experimental results show that this algorithm has higher generalization ability.