Research On The Key Technologies Of Selective Multiple Classifiers Ensemble

In recent years, as the development of compute technology, multiple classifiers ensemble has become a hot topic in both the pattern recognition and machine learning. The necessary and sufficient condition with multiple classifiers ensemble being superior to a single classifier is that the classifiers in the ensemble are not only precise but also wrong diversity. On the beginning ,the method of producing the individual classifiers directly to ensemble which satisfies the above conditions was introduced. But individual classifiers which are directly produced are not necessarily wrong diversity, namely that the mistakes which the different individual classifiers produce in the characteristic space does not necessarily distribute in the different regions. So some individual classifiers are redundant. In the actual pattern recognition system, on the premise of guarantee classification ability, the less the number of the ensemble classifiers is, the lower the complexity of classifiers ensemble system is.So now we mostly use a classifiers ensemble design strategy which is called the "over production, again selection" strategy. In the stage of over production we can produce many base classifiers for ensemble directly. The stage of again selection optimizes the over production stage in order to obtain a classifier subset which classified recognition performance best. The researcher hoped that performance of multiple classifiers ensemble which is forecasted by the kernel cluster and the diversity measure can instruct the selection of multiple classifiers.This article researches to the selective part of classifiers ensemble according to the two conditions which guarantee the multiple classifiers ensemble more effective: the classifiers in ensemble are precise; the classifiers in ensemble are error diversity. The prime tasks of article are as follow:Firstly, a selective multiple classifiers ensemble method based on the diversity measure DMSE is proposed, which introduces Zhou-zhihua's selective ensemble method SEME to the selective multiple classifiers, and uses the diversity measure method to select the ensemble classifiers. The result of experiments indicates that this algorithm reduces the number of the ensemble classifiers, and the accuracy rate is higher than bagging.Secondly, a selective multiple classifiers method based on the kernel cluster KCSE is proposed, which introduces kernel cluster to the selective multiple classifiers. It uses kernel cluster algorithm which is the non-surveillance to divide the entire sample characteristic space into several parts, then select classifiers to compose classifiers subset of entire sample space according to classification accuracy rate of classifier to samples in various cluster characteristic region and the diversity. The theoretical analysis and the result of experiments indicate that the algorithm can obtain good classified effect.Thirdly, a selective multiple classifiers method based on the kernel cluster and kernel cluster KCDMSE is proposed considering the accuracy rate and diversity of base classifiers, which introduces kernel cluster and kernel cluster to the selective multiple classifiers. In the stage of the accuracy rate selection using the kernel cluster, we calculate the recognition rate according to the characteristic that the recognition rate of classifiers is high in some sample region recognition rate but other region low. In the second stage of diversity selection, we select classifiers subset according to the classifiers which are diversity in the ensemble can promote the ensemble. The two selections have guaranteed the base classifiers the recognition accuracy rate high and the wrong diversity, the multiple classifiers ensemble system is simplified and optimized. The experiments on UCI database and ELENA database indicate that the algorithm has a high classified accuracy rate.