Several researchers have shown that multiple classifier systems can result in effective solutions to difficult real-world classification tasks. However, most of these approaches are easily influenced by noise, and the training datasets for local classifiers get easily imbalanced. One of the main reasons for this is that it is hard to guarantee that the centers of the subsets are close to the separation hyperplane, so that it is difficult to evenly distribute the samples in the two sides of the hyperplane. In order to solve this problem, we redefine the description of classifier modeling problem as a task of piecewise approximation of the separation hyperplane. On the basis of this description, we propose a novel multiple support vector machine (SVM) classifier system. Its main contribution is a novel construction approach to the subtraining datasets. The proposed approach partitions the area close to the separation hyperplane into some subsets to construct the subtraining datasets. The subtraining datasets describe the subtasks for identifying segments of the separation hyperplane. Local SVMs are trained to solve the respective subtasks. Finally, the decisions of these local SVMs are appropriately combined on the basis of a probabilistic interpretation to obtain the final classification decision. The effectiveness of this approach is demonstrated through comparisons with some well-known approaches on both synthetic and real-world datasets. © 2013 Institute of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc.