Maximal margin hyper-sphere SVM for binary pattern classification

In this paper, we propose a novel maximal margin hyper-sphere support vector machine (MMHS-SVM) for binary pattern classification. Our proposed MMHS-SVM aims to find two hyper-spheres simultaneously by solving a single quadratic programming problem and is consistent between its predicting and training processes. An essential difference that distinguishes it from other hyper-sphere SVMs is that the optimization model is constructed by maximizing the sum of the square distance between centers of two hyper-spheres, but not the sum of squares distances from the center of hyper-sphere to all examples of the opposite class. Such a principle of structural risk in our MMHS-SVM not only helps us grasp the critical samples and eliminate a large number of redundant samples, but also reduces the test cost due to the sparsity. In addition, an effective SMO-typed algorithm is designed to decrease the high time complexity and storage. Finally, a large number of experiments verify the above statements again. The experimental results on several artificial and publicly available benchmark datasets show the feasibility and effectiveness of the proposed method.