Research On Dynamic Fusion Technology Of Multiple Classifiers

The purpose of multi-classifier is to combine advantages of each base classifier's performance together to predict.So that we could obtain the best classification model.Multi-classifier dynamic fusion technology,whose main idea is to choose different base classifier by different sample to achieve greater flexibility and higher precision.Therefore,multi-classifier fusion technology has become a research hotspot with endless research results.However,there are still some problems in this technology,such as the construction of differences among multiple classifiers,the setting of conditions for the combination of dynamic filter base classifiers based on KNN,and the setting of weights in weighted fusion.In view of above defects,the main research of this paper is as follow:1.Aiming at the scope of application of subsequent dynamic fusion algorithm,this paper purpose a set of date screening rules at first.This rule determines whether all the base classifiers can predict correctly in the capability region by finding the neighbor-sample to form a capacity area.If all the base classifiers could predict correctly their neighbors,the sample doesn't need to be dynamically fused.Otherwise,it builts a subtest set.2.Aiming at the defects of KNORA-ELIMINATE-W method of selecting and intergrating,this paper purpose an improved KNORA-ELIMINATE-W method,which tests capacity of base classifier in this capacity area by finding the neighbor samples of subtest set.If a base classifier could predict these neighbors correctly,the base classifiers is a member of the set.Otherwise,we should use prediction result of the base classifier with highest local accuracy as the classification result.3.In order to further improve the accuracy of classification,this paper purpose a two-layer hybrid algorithm based on Stacking.The first layer model use single-classification algorithm and the second layer uses the improved KNORA-ELIMINATE-W method.Finally,in order to evaluate all the classification performance in this paper.We compared the purposed method with others.The accuracy is improved by about 2%,which proves its effectiveness.