Abstract:  For small data sets, there exists many machine learning algorithms, such as neural networks, naive baycs classifier, decision tree and support vector machine，ctc，can get very good performance. But for largrscalc problem, the performance of these learning algorithms is not satisfactory. Then we always resort to ensemble learning. Min-Max modular support vector machines(M3-SVM) is one of effective ensemble learning methods. This approach has successfully been applied in many fields of pattern classification. One of the key problems of M3- SVM is to find an effective and low-complexity partitioning method of training samples, then to shorten the training time and to get relatively balanced training subsets. The advantages of traditional K-means clustering arc simple and
low time complexity. However, it is sensitive to initial point selection. The criterion function is generally optimized by a gradient method，and the search direction of the gradient is along with the direction of energy decreasing, so the result is often local optimal solution rather than global one. In the paper, a new partitioning method is presented，which based on bisecting K-means. For the bisecting clustering, dichotomy strictly belongs to hierarchy clustering. And hierarchy clustering forms a hierarchical tree structure，which contains the information of all levels and the similarity within and between clusters. So the bisecting K-means algorithm can get a global optimal solution and its time complexity is still low. Furthermore, it can get relatively balanced training subsets by means of hierarchical
clustering.The experimental results on real-world datasets show that this partitioning method can get compromise between the the training time and classification accuracy rate.