树分类器的渐近最优训练算法

An Asymptotically Optimal Tree Classifier Training Algorithm

  • 摘要: 传统的非参数树分类器设计都采用先划分再编辑(即按—区域内各类训练模式的多少把这区域划归多数类的决策域)的方法.这种设计路线限制了识别率的提高.本文给出的先编辑训练样本再划分空间的设计方法,可以使树分类器达到依概率渐近圾优的识别率.相应的计算机仿真试验表明,这种树分类器在模式分布复杂和训练样本容量较小的情况下仍能达到较高的识别率;而且它的时间效率比其他非参数分类器,如近邻式分类器高出100至1000倍.

     

    Abstract: Conventional non-parametric tree classifier design is characterized by partition-edit,where each subregion is assigned to a class according to voting by the patterns in thisregion.This scheme usually does not result in high recognition rate.An edit-partitionscheme is given in this paper,which achieves asymptotically Bayesian optimal recogni-tion rate.Simulation experiments have been conducted,which show that tree classif-iers in this scheme achieve high recognition rate even when the pattern distribution iscomplex and the training sample size is small.Furthermore the time efficiency of edit-partition tree classifiers is 100 to 1000 times higher than that of the other non-param-etric classifier,e.g.the nearest neighbor rules.

     

/

返回文章
返回