采用邻域决策分辨率的特征选择算法  

Feature Selection Algorithm Based on Neighborhood Decision Distinguishing Rate

作　　者：诸文智[1] 司刚全[1] 张彦斌[1] 

机构地区：[1]西安交通大学电气工程学院,西安710049

出　　处：《西安交通大学学报》2013年第2期20-27,共8页Journal of Xi'an Jiaotong University

基　　金：国家自然科学基金资助项目(61005058);教育部高等学校博士学科点专项科研基金资助项目(20090201110019;20070698059)

摘　　要：针对目前基于粗糙集模型的特征选择算法无法直接应用于数值型数据、必须经过离散化过程而造成决策信息丢失的问题,提出了一种基于邻域决策分辨率的特征选择算法。该算法根据邻域信息粒中决策分布与其分类能力间的关系,提出了邻域决策确定性(Nc)来衡量单个信息粒的决策分辨能力;并根据特征向量空间上所有信息粒所具有的Nc累加值,定义了邻域决策分辨率作为特征子集上决策可分辨性的量度,从而将名义型和数值型数据统一在同一特征选择算法框架下。仿真实验和实际应用的结果表明,该算法性能优于目前主流基于邻域粗糙集的特征选择方法。The current feature selection algorithms based on the neighborhood rough set（NRS） model are unable to evaluate numerical dataset directly,a discretization procedure becomes necessary to transform the datasets into discrete forms,but inevitably leads to useful decision information loss.To solve this difficulty,a feature selection algorithm based on the neighborhood effective information rate is proposed.In view point of granulated neighborhood,the relation between the decision discernibility and the decision distribution is analyzed,and the neighborhood decision certainty（Nc） is defined to indicate the degree of distinguishing capability in each individual neighborhood granule.The neighborhood decision distinguishing rate（NDDR） of the feature subset,which evaluates the ability of the subspace to approximate decision space,is established based on the sum of the Nc values of the information granules induced by the corresponding feature space.Then the nominal and numerical datasets can be integrated into the same feature selection algorithm framework.The simulation and application illustrate that the proposed algorithm outperforms the other NRS-based ones.

关 键 词：特征选择 邻域粗糙集 邻域决策分辨率 

分 类 号：TP274[自动化与计算机技术—检测技术与自动化装置]