Selective Ensemble Method Based On Diversity Measures

Classifiers ensemble plays an important role in the development of pattern recognition.As the number of classifiers increases,the diversity between classifiers will be reduced,and then some redundancy will be generated and the accuracy of the ensemble will be affected.Therefore,some scholars put forward strategy of selective classifiers ensemble.By adding a classifier to the selection stage,we remove the classifier with less difference in the initial base set.This paper focuses on two main goals of selective ensemble: the set of selected classifier has appropriate diversity and the selected classifiers have good ensemble performance.The main work of this paper includes:(1)A selective ensemble method based on clustering and sorting pruning is proposed.Firstly,the confusion matrix of the classifier set is converted to cluster,and then based on the DREP(diversity regularized ensemble pruning)algorithm proposed by Zhou Zhihua et al.,a sort pruning algorithm is proposed for the parallel pruning of each subset.Finally,the classifier ensemble is integrated by voting method.The experiment is carried out on 10 sets of data sets in the UCI database and compared with the Bagging and DREP algorithms.The results show that the algorithm improves the classification ability of the integrated system effectively.(2)A classifier selection method using mixed diversity measures is proposed.First,we calculate the pairwise difference matrix of the classifier set,then convert it into the adjacency matrix of the graph according to the threshold,and then use the graph coloring method based on the genetic algorithm to group the classifier.Finally,a kind of evaluation system based on information entropy is proposed.The evaluation index is a number of non-pairwise difference measurement methods,and the best group to participate in the final integration is selected according to the system.The experiment is compared with a variety of integrated methods(Bagging,Adaboost,DREP,Diversity Measures for Ensemble Pruning(DivP)),and the results prove the feasibility of the algorithm.