The Research For Multiple Classifier Dynamic Ensemble And The Application

A good machine learning system must have a strong generalization. That's, the model built according to the available data can handle the new data very well. In order to achieve good generalization ability and robustness, multiple classifiers ensemble methods integrate individual many devices (base classifiers) with certain rules, which makes full use of complementarity among base classifiers.As one of important branchs of multiple classifier ensemble, multiple classifier dynamic ensemble integrates different base classifiers according to diffierent samples, so it has good flexibility and effectiveness and it is more effective than traditional method. Because of these, more and more pelple are doing research in the field. In order to get strong generalization ability, this paper does systematic study of multiple classifier ensemble and multiple classifier dynamic ensemble. The main work of this paper includes:(1) The paper gives a general introduction about research background, significance and current status of multiple classifier ensemble and gives a introduction about the priciple in detail. After the overview of the framework of multiple classifier ensemble method and dynamic classifier ensemble method, this paper makes an analysis of the reason why they are useful, then does a comparison among them and finally makes an analysis of their limitation.(2) Proposed a dynamic ensemble model based on weighted error rate. Through making analysis of limitation of popular dynamic classifier ensemble methods, we find that the popular methods take no account of samples outside of the capacity region and treat samples inside equal, and most of them have to set parameters manually. That's why the popular methods'performance is instability. According to these, we make the entire sample set as the capacity region and treat the samples diffierent. The experiment results in UCI data sets and features of BMP steganalysis, compared to the popular ensemble methods, the dynamic ensemble metods based on weighted error rate have good performance with dynamic selection, dynamic weighted ensemble and selective_weighted ensemble.(3) Proposed a new steganalysis algorithm based on dynamic random subspace and Fisher (DCSW_RF). The traditional method in steganalysis is like this:First, features must be extracted by some extracting algorithms; second, a SVM model can be get with the extracted features; at last, new samples can be tested by the SVM model. But, with the increase in the number of feature dimensions, the training speed and the final results sometimes cannot be satisfying. In this article, I use DCSW_RF instead of SVM. The result shows that DCSW_RF is faster and more accurate.