基于二次互信息的特征选择算法

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.

Feature selection based on quadratic mutual information criterion

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.