基于 Parzen 窗条件互信息计算的特征选择方法

摘要

In order to solve the problems of calculating the conditional mutual information of continuous variables and bias of multi-value features,this paper proposed a novel feature selection method.The method was based on computing conditional mutual information with Parzen window called PCMIFS,which adopted Parzen window to estimate the probability density func-tion and compute conditional mutual information of continuous feature.And introduced a penalty factor,feature dispersion,to overcome the bias of multi-value features.The experiment results show that comparing several existing method,PCMIFS can attain better or comparable performance,and is an effective feature selection method.%为解决连续值特征条件互信息计算困难和对多值特征偏倚的问题，提出了一种基于 Parzen 窗条件互信息计算的特征选择方法。该方法通过 Parzen 窗估计出连续值特征的概率密度函数，进而方便准确地计算出条件互信息；同时在评价准则中引入特征离散度作为惩罚因子，克服了条件互信息计算对于多值特征的偏倚，实现了对连续型数据的特征选择。实验证明，该方法能够达到与现有方法相当甚至更好的效果，是一种有效的特征选择方法。