Clustering analysis application in SVM ensemble^#

Aiming at the existing support vector machine ensemble's problems, e.g. strong randomicity, larger scale of the training subsets and high complexity of the ensemble classifier, this paper put forwards a selective support vector machine ensemble method based on clustering analysis. Firstly, the samples are clustered into several clusters with rival penalty competitive learning algorithm. Then the instances from different positive and negative clustering subset are respectively chosen to create support vector machine components. Finally the most appropriate base classifier is selected to classify the new examples according to the distance between the testing point and the cluster centers in the input space. Experiment results on synthetic and UCI datasets show that the support vector machine ensemble generated by our method has higher classification accuracy and much lower time and space complexity than Bagging, Adaboost and k-fold cross validation algorithms.