A support vector machine （SVM） ensemble classifier is proposed. Performance of SVM trained in an input space eonsisting of all the information from many sources is not always good. The strategy that the original inp...A support vector machine （SVM） ensemble classifier is proposed. Performance of SVM trained in an input space eonsisting of all the information from many sources is not always good. The strategy that the original input space is partitioned into several input subspaces usually works for improving the performance. Different from conventional partition methods, the partition method used in this paper, rough sets theory based attribute reduction, allows the input subspaces partially overlapped. These input subspaces can offer complementary information about hidden data patterns. In every subspace, an SVM sub-classifier is learned. With the information fusion techniques, those SVM sub-classifiers with better performance are selected and combined to construct an SVM ensemble. The proposed method is applied to decision-making of medical diagnosis. Comparison of performance between our method and several other popular ensemble methods is done. Experimental results demonstrate that our proposed approach can make full use of the information contained in data and improve the decision-making performance.展开更多
基金Supported by the High Technology Research and Development Programme of China （2002AA412010）, and the National Key Basic Research and Development Program of China （2002cb312200） and the National Natural Science Foundation of China （60174038）.
文摘A support vector machine （SVM） ensemble classifier is proposed. Performance of SVM trained in an input space eonsisting of all the information from many sources is not always good. The strategy that the original input space is partitioned into several input subspaces usually works for improving the performance. Different from conventional partition methods, the partition method used in this paper, rough sets theory based attribute reduction, allows the input subspaces partially overlapped. These input subspaces can offer complementary information about hidden data patterns. In every subspace, an SVM sub-classifier is learned. With the information fusion techniques, those SVM sub-classifiers with better performance are selected and combined to construct an SVM ensemble. The proposed method is applied to decision-making of medical diagnosis. Comparison of performance between our method and several other popular ensemble methods is done. Experimental results demonstrate that our proposed approach can make full use of the information contained in data and improve the decision-making performance.
