Research On Dynamic Ensemble Learning Algorithm

Ensemble learning can significantly improve the accuracy of classification and has always been a research hotspot in the field of machine learning.Traditional ensemble learning algorithms failed to consider the local validity of the base classifier,did not take into account the complexity of large data sets,and failed to evaluate and select the classifiers participating in the integration.The structure is not flexible enough,resulting in the face of large data sets,classification accuracy is low and inefficient,so more effective ensemble learning algorithms such as selective ensemble and dynamic multi-classifier integration techniques are proposed.This paper mainly focuses on the dynamic multi-classifier integration technology,and studies two dynamic integration algorithms: one is focused on the construction method of the classifier,the base classifier is generated in different subsets,and the ensemble learning algorithm is integrated by using the decision tree.The other is a dynamic fusion method—a dynamic integration algorithm based on correlation analysis.The main research content can be divided into two parts.In the first part,based on the idea of gradient optimization,an ensemble learning system with decision tree structure is proposed.The ensemble algorithm proposed in this paper uses classifier classification categories to creating smaller subsets for next level classifiers,use these smaller subset of training samples to construct base classifiers with better local classification accuracy.An ensemble learning system that has a decision tree structure is generated through the integration of the base classifier.An experiment was performed on a real data set of American College Matriculation Set.The experimental results verify that the algorithm can effectively reduce the variance of the ensemble system while keeping the classifier bias unchanged and improve generalization performance.In the second part,from the perspective of classifier fusion,a dynamic fusion method based on correlation analysis is proposed.Many dynamic ensemble algorithms determine the effective classification region of the classifier according to the similarity measure between the measured sample and the training sample,so as to achieve better classifier dynamic fusion.In this paper,the learning method is used to obtain more precise classification relevance.A selection layer is added to the upper layer of the multi-classifier system used to dynamic select base classifiers.The implementation of selection behavior of selection layer is achieved by training which use marking information on whether the classification is correct or not on multiple classifiers.At the same time,in order to improve the integration effect,the base classifier is firstly screened and sorted by the complementary index and the correct classification coverage.Experiments were conducted on multiple experimental data sets.The results show that the proposed algorithm is effective.This paper studies how to improve the performance of the integrated classification system from the two different aspects of classifier construction and integration,and has done a lot of experiments on the actual data set,which provides some feasible solutions for solving practical problems...