一种新的Boosting回归树方法

A New Boosting Regression Tree Method

梯度Boosting思想在解释Boosting算法的运行机制时基于基学习器张成的空间为连续泛函空间,但是实际上在有限样本条件下形成的基学习器空间不一定是连续的。针对这一问题,从可加模型的角度出发,基于平方损失,提出一种重抽样提升回归树的新方法。该方法是一种加权的加法模型的逐步更新算法。实验结果表明,这种方法可以显著地提升一棵回归树的效果,减小预测误差,并且能得到比L2Boost算法更低的预测误差。

The basis of gradient boosting idea aimed to explain the working of boosting is that the space spaned by base learner is continuous functional space. But in practice,this space is not continuous under limited sample.To this problem,under the point of additive model view,in this study,a new resample boosting regression tree algorithm is proposed.This algorithm is a stage wise method in resample additive model.Our numerical experiments demonstrate the algorithm can improve results of a regression tree,reduce prediction errors evidently and get lower prediction error than L2Boost algorithm.