基于加权最大边缘间距准则MMC的特征选择问题

特征选择是模式识别经典而重要的课题．由于不同类别样本之间存在边缘样本点，其分布区域互相交叉重叠，经典的MMC（Maximize Marginal Criterion）方法简单地采用最大化类中心距离，不利于样本分类．针对此问题，给出了一种基于加权最大边缘间距准则（加权MMC）并改进了的特征选择算法，该方法考虑了不同类别数据边缘样本点在模式分类中的作用，建立了基于最大边缘间距的新型特征评分准则，提高了边缘样本点在衡量特征判别能力时的作用．在公开数据集PIE和MIT-CBCL3000标准人脸图像库上进行了实验，结果表明，该算法与经典的MMC特征选择算法相比较具有明显的优势．

On Feature Selection of Weighted MMC Distance Based on the Maximum Margin Criterion

Feature selection is a classical and important subject of pattern recognition. Due to the edge points among different samples whose regional distributions are overlapping, the classical MMC method is not conducive to the sample classification by simply using the maximum center distance. Therefore, a feature selection algorithm based on weighted Maximize Marginal Criterion （ MMC ） is proposed, considering the roles of edge sample points of different types of data in the pattern classification, establishing a new feature score criterion based on maximum edge distance, and improving the functions of marginal sample points in measuring feature discrimination ability. The experiments on public data set PIE and MIT-CBCL3000 face image database show that the proposed feature selection algorithm in this paper has obvious advantages compared with the classic MMC method.