The Research And Application On The Ensemble Algorithm Based On Radial Basis Neural Network

The types of datasets which need to be handled increase rapidly that cause the accuracy of the data processing and the stability of algorithm become more and more important. The single classifier could not satisfy the request yet, so the ensemble algorithm catches the attention of researchers, the decision tree algorithm is one of most common algorithm used in the ensemble algorithm; however, the basis classifier in the decision tree is always composed of weak classifiers which will engender the over-fitting problem. So, this is one of the problems we need tackled on the balance of precision improved and the over-fitting phenomenon decreased.For the sake of handling binary classification datasets, the usual algorithms always give a weak precision, so, we take the Receiver Operator Characteristic (ROC) which under 95% degree of confidence to reduce the datasets redundancy. Then the datasets which has been processed by Synthetic Minority Over-sampling Technique (SMOTE) are used by the hybrid of random forest and RBFNN algorithm to improve the over-fitting problem as the precision increased. At the same time, the Diversity Ensemble Creation by Oppositional Relabeling of Artificial Training (DECORATE) Examples algorithm is used to remodeling the datasets to handle multi-label classification datasets. We take the technology of interpolation among identically distribute, then we use the RBFNN as the basis classifier of DECORATE. It will give a higher precision as the adaption of different kinds of datasets increased. In the purpose of increasing the diversity among the basis classifiers, the algorithm of Rotation Forest with RBFNN ensemble algorithm is proposed. The ensemble algorithm departs the input features into lots of small ones. Then the PCA is used on these new small features which can get the PAC coefficient. After the change of datasets, we remodel the Rotation Forest by RBFNN as the basis classifier to improve the over-fitting phenomenon. These three algorithms are used in the UCI standard datasets and the remote sensing datasets, the final show that these algorithms have the availability of higher classification result and good generalization level.