The Research On Neural Network Ensemble And Application In The Classification Of Cancer

The expression level of thousands of genes in cell can be simultaneously observed in different conditions by micro-array experiment. And the method how to analyze the gene expression data is the hot issue in bioinformatics research. In the meanwhile the method has great effects on the research in the field of biomedicine. In the research of cancer, better way to cancer's diagnosis and therapy will be produced via studying on the difference in cancer gene expression data, which comes from micro-array technique.Artificial neural network has been applied in many fields. The technique of neural network ensemble, through the ensemble of many individual neural networks for enhancing the system's generalization, is the hot topic in the field of neural computing. In this paper, we studied the neural network ensemble's application in the cancer's classification.For the small samples in the gene expression data, Bagging and Boosting are introduced to analyse the algorithm of neural network ensemble. Boosting set the sample, which is hard to be classified, more chance to be studied by neural network, that means give this kind of sample a high probability for sampling to improve the result of classification. But it is unstable. Bagging is the method with selecting sample by the same probability for training data. The result with Bagging is not as good as the result with Boosting, but it is stable. A new method with sampling by different probability, for the training data to train the individual network, and was performed in the gene expression data. The results shown the method enhance the ability of classification. In the analysis of the gene expression data, a proved method of signal to noise was put forward for the selection of character gene, which is using average value instead of median data, and was performed in the gene expression data. And the result of experiment showed the method removes the redundant genes effectively.For the selection of character gene, a feature selection method based on the stratified sampling was proposed for choosing gene as the member of character gene group, and experiment was implemented in the colon tumor dataset .The results had shown character gene group chosen by this method is more effective and rational. Every individual network for the ensemble is chosen by the accuracy of classification. In experiment this method improved the accuracy of classification.