The Research On Dynamic Ensemble Classifiers

The fundamental issue for classification problem caring about is how to effectively improve the generalization ability of the classification system. Though traditional classification techniques have been successfully employed in some practical application fields, and the classification performances obtained by them are recognized to some degree, however, people put forward the higher requirements towards the technological achievements. This implies that it needs to develop new effective techniques to meet the demands of people with the progress of the times and the development of science and technology. The ensemble of classifiers is produced under such background. It is desirable to combine the predictions of multiple learners to improve the generalization ability and to strengthen the robustness for the classification system. The ensemble of classifiers has become a hot research topic in machine learning, data mining, and other domains. Now, the relevant research area gathers more and more experts and scholars and appears more and more study fruits accordingly.This paper comprehensively reviews abroad and domestic research status, basic principle and typical methods of the ensemble of classifiers. The emphasis is placed on the dynamic ensemble technique which is an important branch of ensemble learning. Moreover, the main shortcomings about the study of the dynamic ensemble mechanism are summarized while the corresponding solutions are presented as well in order to improve the classification performance of the integrated system.The research fruits about this paper are described as follows:1. The developing status, related concepts and working mechanism of the ensemble of classifiers are systematically introduced. Then three classical ensemble methods are discussed in details.2. Basic principle and framework of dynamic ensemble mechanism are concluded. Typical dynamic ensemble approaches are detailedly talked about including their main disadvantages.3. In order to reduce the risk brought by the shortage of training samples which would lead to the unreliableness of estimating local performances of base classifiers, four new dynamic weighting ensemble methods, called DWEC-CV-KOLP, DWEC-CV-KLCP, DWEC-CV-OLA, DWEC-CV-LCA, are proposed which introduce cross-validation technique to dynamically assign a weight to each component classifier. Experiments on several UCI data sets show that when the size of training set is not large enough, the proposed methods can achieve better performances compared with some classical ensemble approaches.4. For the sake of eliminating the false neighbors in the neighborhood of the test sample, which can interferes with the estimation of local performances of base classifier, an improved dynamic weighting ensemble scheme called DWEC-CV-MCB is presented which injects multiple classifier behavior information into the evaluation process. Experimental results demonstrate that the improved approach can not only achieve better classification performances, but also lower the risk and cost brought by setting up parameter K manually.5. The issue that how to enhance incremental learning effect through utilizing dynamic ensemble mechanism is investigated. For this purpose, an incremental classification model based on dynamic selection is put forward. Through improving two shortages of Learn++, a typical incremental classification method, the raised model gets better classification results.