基于一致性和知识粒度的半监督特征选择方法  

Semi-supervised Feature Selection Method Based on Consistency and Knowledge Granularity

作　　者：万丽娟 钱文彬[1] 曾武序 WAN Lijuan;QIAN Wenbin;ZENG Wuxu(School of Software,Jiangxi Agricultural University,Nanchang 330045,China;School of Computer and Information Engineering,Jiangxi Agricultural University,Nanchang 330045,China)

机构地区：[1]江西农业大学软件学院,江西南昌330045 [2]江西农业大学计算机与信息工程学院,江西南昌330045

出　　处：《山西大学学报（自然科学版）》2023年第1期53-61,共9页Journal of Shanxi University(Natural Science Edition)

基　　金：国家自然科学基金(61966016);国家重点研发计划项目(2020YFD1100605);江西省自然科学基金(20224BAB202020);江西省研究生创新专项基金项目(YC2020-S236)。

摘　　要：针对数据标注的代价昂贵和半监督学习难以直接处理高维数据,其包含的冗余特征往往导致分类模型效果不理想问题。为了解决上述问题,根据粒计算模型,提出基于一致性和知识粒度的半监督特征选择方法。利用正域的依赖度去度量有标记样本的一致性,同时采用知识粒度对未标记样本去评价特征对样本空间的可区分性,由此结合数据分布情况构造了一种基于线性融合的特征重要性方法。在此基础上,设计了面向半监督数据的特征选择方法。最后,通过实例分析和与当前四种半监督特征选择方法对比进一步验证了本文方法在半监督数据中的有效性和可行性。The major problems in data mining are the high cost of data labeling, and the difficulty for semi-supervised learning to directly deal with high-dimensional data, which can not result in ideal classification performance due to redundant features. In order to solve the above problems, according to the perspective of granular computing theory, a semi-supervised feature selection method is proposed by combining consistency and knowledge granularity. In the proposed framework, the dependence of the positive region is used to measure the consistency of the labeled instances, and the knowledge granularity is used to measure the distinguishability between the feature to the sample space of the unlabeled data. Then, combining with the data distribution, a measure of feature importance based on linear fusion is proposed. Based on this a feature selection method for semi-supervised data is designed. Finally,by the example analysis and experimental comparison, the superiority and feasibility of our proposed method is further demonstrated on the semi-supervised data. The proposed algorithm can obtain feature reduction results with better classification performance.

关 键 词：特征选择 粗糙集 一致性 知识粒度 半监督数据 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]