A Max Modular Support Vector Machine and Its Variations for Pattern Classification

In this paper, we propose a max modular support vector machine (M²-SVM) and its two variations for pattern classification. The basic idea behind these methods is to decompose training samples of one class into several parts and learn each part by one modular classifier independently. To implement these methods, a dasiapart-against-otherspsila training strategy and a max modular combination principle are proposed. Also, a down-sampling technique is employed to improve the training without losing any information of positive sample. Finally, a variation called feature-constructed M²-SVM (FM²-SVM) is proposed,which selects features of each subset for classification in one modular. Experimental results show that FM²-SVM can not only improve the precision of classifying but also reduce the dimension of input features. Performances of the proposed methods are shown to be superior to traditional SVMs as well as M³-SVM and KNN on artificial data, UCI Forest Cover type data, CASPEAL face database and AR face database.