Research On Classifier Selective Ensemble Method And Thire Diversity Measurement

Ensemble learning is one of the focuses of machine learning and it combines a number of different models into one single model which aims to use the difference between these individual models to improve the generalization performance of the model. That is designed to give full play to each member of the classifiers in classification performance and gets higher classification rate than individual member classifier.In recent years, classifier selective ensemble technology has been widely applied in speech recognition, face recognition, image recognition,data mining, medical diagnosis, games, handwritten character recognition, remote sensing image classification and many fields and it shows a great deal of research value and realistic prospects.The purpose of Classifier ensemble is to use different complementarity between classifiers to improve classification performance after integration.Usually, by increasing the recognition performance of member classifier and increasing the difference of member classifiers to achieve the aim of improving performance of classifier ensemble. The traditional method of classifier ensemble has some deficiency to show the diversity. In order to achieve the optimum performance of combined classifier, it needs to generate a number of classifiers with high accuracy and difference, and also need to consider different types of samples for the integration of different ways.This paper proposes the optimization of two-stage (The generation stage and the combination stage of the individual classifier) about ensemble learning and have optimization from the generated stage to assembly phase. In the generation phase of classifier, in order to generate higher classification accuracy, we use the way of feature segmentation for original training data set; In the stage of combination of classifier, we use a reasonable difference metric formula to select the classifier which has a difference, so these two needs are guaranteed and balanced, it will improve the recognition performance of ensemble learning.The aim is to achieve the purpose of the diversity of member classifiers in case of ensuring high-performance of single classifier.