基于out-of-bag样本的随机森林算法的超参数估计

摘要

Random forest (RF) is an effective decision tree ensemble method. In order to achieve its best performance, however, the optimal value of the hyper-parameter in RF needs to be estimated by an appropriate method. Under the condition that the computational cost is not additionally consumed, this paper proposes a new approach to estimate the hyper-parameter based on the out-of-bag sample. The experiments conducted by some UCI real-world data sets show that RF with the hyper-parameter estimated by the proposed method performs best in most cases.%随机森林是一种有效的分类树集成算法,但为了使它具有较高的预测精度,要采用某种方法确定其超参数的最优值.在不额外增加计算复杂性的前提下,提出了一种基于out-of-bag样本估计其超参数取值的方法.仿真试验的结果表明,利用文中提出的方法所选取的超参数在多数情况下都能使随机森林算法的分类效果达到最优.