改进的文本特征选取算法研究  

Research on Improved Text Feature Selection Algorithm

作　　者：朱世玲 郑彦[1] ZHU Shi-ling;ZHENG Yan(School of Computer Software,Nanjing University of Posts and Telecommunications,Nanjing 210023,China)

机构地区：[1]南京邮电大学计算机软件学院,江苏南京210023

出　　处：《计算机技术与发展》2019年第5期66-69,共4页Computer Technology and Development

基　　金：国家"863"高技术发展计划项目(2006AA01Z201)

摘　　要：特征选取的好坏决定了文本分类的准确度。文本特征选取通常有文档频率、互信息、信息增益、卡方统计量等方法。文中讨论了文档频率和互信息在特征选取时的缺点,基于这些缺点,提出了一种混合文档频率和互信息的改进算法。文档频率进行特征选取时会偏向选择高频词,而没有考虑到该词是否在类别间有区分度,所以提出通过计算词的文档频率的类别方差作为文档频率的权重来进行特征选取。互信息偏向选择低频词,也忽略了互信息值为负的那些特征作用,有些互信息为负的词反而包含更多的类别信息。所以对互信息的值取了绝对值来加强互信息为负的词的作用。通过对比DF、MI和改进的DFMI的实验结果,发现该算法在精度、召回率和F_1度量上都有所提高,验证了该方法的有效性。The quality of feature selection determines the accuracy of text classification.Text feature selection usually includes document frequency,mutual information,information gain,Chi-square statistics and so on.We discuss the shortcomings of document frequency and mutual information in feature selection,and on the basis propose an improved algorithm for hybrid document frequency and mutual information.In the feature selection of document frequency,high-frequency words are preferred without considering whether there is a degree of discrimination between categories.Therefore,we propose to take the category variance of word frequency as the weight of document frequency for feature selection.The mutual information tends to select low-frequency words,but also ignores those features with negative mutual information value.Some words with negative mutual information contain more category information.Therefore,the absolute value of mutual information is taken to strengthen the role of words with negative mutual information.The experimental comparison of DF,MI and improved DFMI indicates that the proposed algorithm improves in accuracy,recall rate and F 1 measure,which verifies its effectiveness.

关 键 词：特征选取 互信息 文档频率 文本分类 改进互信息 改进文档频率 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]