基于二次互信息的特征选择算法  

Feature selection based on quadratic mutual information criterion

作　　者：李滔[1] 王俊普[1] 王鹏伟[2] 吴秀清[2] 

机构地区：[1]中国科学技术大学自动化系,安徽合肥230027 [2]中国科学技术大学电子工程与信息科学系,安徽合肥230027

出　　处：《中国科学技术大学学报》2006年第11期1133-1138,共6页JUSTC

基　　金：中国高技术研究发展(863)计划项目(2002AA783055);中国科学院知识创新项目(KZCX0101)资助

摘　　要：将二次互信息(mutual information)用作模式分类问题中特征选择的准则,分析了该准则在再生核希尔伯特空间中的几何意义.在二次互信息准则基础上,提出了基于Parzen窗密度估计和后向删除策略的特征选择算法PW-QMI,同时针对大规模数据集的情况给出了基于高斯混合模型的算法GMM-QMI,以减小算法的计算复杂度.通过与相关度算法和SVM-RFE算法的实验比较,证明了该算法在特征选择问题上具有更为稳定的性能.Quadratic mutual information was used as recognition problems, and the geometric meaning of the criterion for feature variable selection in pattern QMI criterion in the reproducing kernel Hilbert space was analyzed. Based on the criterion, a new feature selection algorithm, PWQMI, was proposed, which used Parzen window for probability density estimation and backward elimination strategy for searching in feature variable space. In situations of a large size example set, another algorithm, GMM-QMI, was proposed which used the Gaussian mixture model for density estimation to reduce computation complexity. The comparative experiments on the correlation-criterion-based algorithm and SVMRFE algorithm show the stable performance of the proposed algorithms for feature selection.

关 键 词：二次互信息 特征变量选择 Parzen窗密度估计 高斯混合模型 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]