At present, studies on training algorithms for support vector machines (SVM) are important issues in the field of machine learning. It is a challenging task to improve the efficiency of the algorithm without reducing the generalization performance of SVM. To face this challenge, a new SVM training algorithm based on the set segmentation and k-means clustering is presented in this paper. The new idea is to divide all the original training data into many subsets, followed by clustering each subset using k-means clustering and finally train SVM using the new data set obtained from clustering centroids. Considering that the decomposition algorithm such as SVMlight is one of the major methods for solving support vector machines, the SVMlight is used in our experiments. Simulations on different types of problems show that the proposed method can solve efficiently not only large linear classification problems but also large nonlinear ones.