资源描述：
《基于区分类别能力的高性能特征选择方法_徐燕》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。
1、ISSN1000-9825,CODENRUXUEWE-mail:jos@iscas.ac.cnJournalofSoftware,Vol.19,No.1,January2008,pp.82−89http://www.jos.org.cnDOI:10.3724/SP.J.1001.2008.00082Tel/Fax:+86-10-62562563©2008byJournalofSoftware.Allrightsreserved.∗基于区分类别能力的高性能特征选择方法1,2+111,2徐燕,李锦涛,王斌,孙春明1(中国科学院计算技术研究所,北京100080)2(华北电力大学,北京102206)A
2、CategoryResolvePower-BasedFeatureSelectionMethod1,2+111,2XUYan,LIJin-Tao,WANGBin,SUNChun-Ming1(InstituteofComputingTechnology,TheChineseAcademyofSciences,Beijing100080,China)2(NorthChinaElectricPowerUniversity,Beijing102206,China)+Correspondingauthor:Phn:+86-10-82522199,Fax:+86-10-62600602,E-mail:xu
3、yan@ict.ac.cnXuY,LiJT,WangB,SunCM.Acategoryresolvepower-basedfeatureselectionmethod.JournalofSoftware,2008,19(1):82−89.http://www.jos.org.cn/1000-9825/19/82.htmAbstract:OneofthemostimportantissuesinTextCategorization(TC)isFeatureSelection(FS).ManyFSmethodshavebeenputforwardandwidelyusedinTCfield,suc
4、hasInformationGain(IG),DocumentFrequency(DF)thresholding,MutualInformation(MI)andsoon.EmpiricalstudiesshowthatIGisoneofthemosteffectivemethods,DFperformssimilarly,incontrast,andMIhadrelativelypoorperformance.OnebasicresearchquestioniswhytheseFSmethodscausedifferentperformance.Manyexistingworkanswers
5、thisquestionbasedonempiricalstudies.ThispaperpresentsaformalstudyofFSbasedoncategoryresolvepower.First,twodesirableconstraintsthatanyreasonableFSfunctionshouldsatisfyaredefined,thenauniversalmethodfordevelopingFSfunctionsispresented,andanewFSfunctionKGusingthismethodisdeveloped.AnalysisshowsthatIGan
6、dKG(knowledgegain)satisfythisuniversalmethod.ExperimentsonReuters-21578collection,NewsGroupcollectionandOHSUMEDcollectionshowthatKGandIGgetthebestperformance,evenKGperformsbetterthantheIGmethodintwocollections.Theseexperimentsimplythattheuniversalmethodisveryeffectiveandgivesaformalevaluationcriteri
7、onforFSmethod.Keywords:featureselection;textcategorization;informationretrieval摘要:特征选择在文本分类中起着重要作用.文档频率(documentfrequency,简称DF)、信息增益(informationgain,简称IG)和互信息(mutualinformation,简称MI)等特征选择方法在文本分类中广泛应用.
