Study On Gradient Boosting Decision Tree And Its Improvement

As is known to all,the regression decision tree model has a clear and intuitive interpretation of the results.This makes it popular in application.But it often can?t provide enough accuracy and result in biggish localization in application.The regression decision tree combined with gradient boosting algorithm can maintain the advantages of the regression tree model,but the improvement of it is finite.In this thesis,we put some tentative improvement measures forward which focus on gradient boosting regression trees algorithm?s localization:(1)we make K-nearestneighbor weighted mean function take the place of simple mean function;(2)we do random subsample before each tree?s training;(3)we change its? old shrinkage step setting to a new shrinkage step which can change itself by learning.We give a hybrid algorithm combined with our improved gradient boosting regression trees and support vector machine and we call it ?hybrid model? for short.In our hybrid model,we not only retain the regression decision tree?s advantages but also enhance our model?s forecast accuracy.Meanwhile,we design some new details on hybrid model?s execution.We make massive data trials on fictitious data sets and nine real data sets to validate the effectiveness of hybrid model.At last,to Enterprise?s stock value assessment problem of private equity investment,we apply our hybrid model to it and give more accurate and more objective analysis with our hybrid model.Compared with the result from traditional enterprise valuation theory,we find our result from hybrid model is more accurate.It indicates our hybrid model is more suitable for enterprise value assessment in actual problems.