A novel Support Vector Machine(SVM) ensemble approach using clustering analysis is proposed. Firstly,the positive and negative training examples are clustered through subtractive clus-tering algorithm respectively. Th...A novel Support Vector Machine(SVM) ensemble approach using clustering analysis is proposed. Firstly,the positive and negative training examples are clustered through subtractive clus-tering algorithm respectively. Then some representative examples are chosen from each of them to construct SVM components. At last,the outputs of the individual classifiers are fused through ma-jority voting method to obtain the final decision. Comparisons of performance between the proposed method and other popular ensemble approaches,such as Bagging,Adaboost and k.-fold cross valida-tion,are carried out on synthetic and UCI datasets. The experimental results show that our method has higher classification accuracy since the example distribution information is considered during en-semble through clustering analysis. It further indicates that our method needs a much smaller size of training subsets than Bagging and Adaboost to obtain satisfactory classification accuracy.展开更多
基金the National Natural Science Foundation of China (No.60472072)the Specialized Research Foundation for the Doctoral Program of Higher Educa-tion of China (No.20040699034).
文摘A novel Support Vector Machine(SVM) ensemble approach using clustering analysis is proposed. Firstly,the positive and negative training examples are clustered through subtractive clus-tering algorithm respectively. Then some representative examples are chosen from each of them to construct SVM components. At last,the outputs of the individual classifiers are fused through ma-jority voting method to obtain the final decision. Comparisons of performance between the proposed method and other popular ensemble approaches,such as Bagging,Adaboost and k.-fold cross valida-tion,are carried out on synthetic and UCI datasets. The experimental results show that our method has higher classification accuracy since the example distribution information is considered during en-semble through clustering analysis. It further indicates that our method needs a much smaller size of training subsets than Bagging and Adaboost to obtain satisfactory classification accuracy.
