基于类别方差的特征权重算法  

Feature weighting algorithm based on class variance

作　　者：周鹏程 刘旭敏[1] 徐维祥[2] Zhou Pengcheng;Liu Xumin;Xu Weixiang(College of Information Engineering,Capital Normal University,Beijing 100048,China;College of Traffic ＆ Transportation,Beijing Jiaotong University,Beijing 100044,China)

机构地区：[1]首都师范大学信息工程学院,北京100048 [2]北京交通大学交通运输学院,北京100044

出　　处：《计算机应用研究》2018年第12期3538-3540,3555,共4页Application Research of Computers

基　　金：国家自然科学基金资助项目(61672002);北京市长城学者资助项目(CIT&TCD20170322)

摘　　要：传统基于概率的特征权重算法,往往只对词频、逆文档频和逆类频等进行统计,忽略了类别之间的相互关系。而对于多分类问题,类别之间的关系对统计又有重要意义。为了提高文本分类的精确度,提出了基于类别方差的特征权重算法,通过计算类别方差来度量类别之间的联系。通过五种特征权重算法在搜狗新闻数据集上的实验,结果表明提出的算法在F1宏平均和F1微平均上都有较大的提高。通过实验验证,该算法提升了文本分类的效果。Feature weight calculation is the basis of text categorization and plays a key role in classification results. Traditional probability based feature weighting algorithms often only count word frequency,inverse document frequency and inverse class frequency,ignoring the relation among classes. For multi classification problems,the relationship between classes is important to statistics. Therefore,in order to improve the accuracy of text categorization,this paper proposed a feature weighting algorithm based on class variance,which calculated the relation between categories by calculating class variance. The experiments of the five feature weighting algorithms in the Sogou news data set show that the proposed algorithm is greatly improved on both the F1 macro mean and the F1 micro mean. Experimental results show that the proposed algorithm improves the effect of text categorization.

关 键 词：特征权重 类别方差 文本分类 支持向量机 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]