欢迎来到天天文库
浏览记录
ID:4124903
大小:618.92 KB
页数:11页
时间:2017-11-29
《基于互信息的无监督特征选择》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库。
1、计算机研究与发展ISSN1000-1239?CN11-1777?TPJournalofComputerResearchandDevelopment49(2):372-382,2012基于互信息的无监督特征选择徐峻岭1,2周毓明2,3陈林2,3徐宝文2,31(东南大学计算机科学与工程学院南京210096)2(计算机软件新技术国家重点实验室(南京大学)南京210093)3(南京大学计算机科学与技术系南京210093)(junlingxu@gmail.com)AnUnsupervisedFeatureSelectionApproachBasedonMutualInfo
2、rmation1,2,ZhouYuming2,3,ChenLin2,3,andXuBaowen2,3XuJunling1(SchoolofComputerScienceandEngineering,SoutheastUniversity,Nanjing210096)2(StateKeyLaboratoryforNovelSoftwareTechnology(NanjingUniversity),Nanjing210093)3(DepartmentofComputerScienceandTechnology,NanjingUniversity,Nanjing210
3、093)AbstractIndataanalysis,featureselectioncanbeusedtoreducetheredundancyoffeatures,improvethecomprehensibilityofmodels,andidentifythehiddenstructuresinhigh-dimensionaldata.Inthispaper,weproposeanovelunsupervisedfeatureselectionapproachbasedonmutualinformationcalledUFS-MI.InUFS-MI,we
4、useafeatureselectioncriterion,UmRMR,toevaluatetheimportanceofeachfeature,whichtakesintoaccountbothrelevanceandredundancy.Therelevanceandredundancyrespectivelyusemutualinformationtomeasurethedependenceoffeaturesonthelatentclassandthedependencebetweenfeatures.Inthenewalgorithm,features
5、areselectedorrankedinastepwiseway,oneatatime,byestimatingthecapabilityofeachspecifiedcandidatefeaturetodecreasetheuncertaintyofotherfeatures(i.e.thecapabilityofretainingtheinformationcontainedinotherfeatures).TheeffectivenessofUFS-MIisconfirmedbythetheoreticalproofwhichshowsitcansele
6、ctfeatureshighlycorrelatedwiththelatentclass.AnempiricalcomparisonbetweenUFS-MIandseveraltraditionalfeatureselectionmethodsarealsoconductedonsomepopulardatasetsandtheresultsshowthatUFS-MIcanattainbetterorcomparableperformanceanditisapplicabletobothnumericalandnon-numericalfeatures.Ke
7、ywordsfeatureselection;unsupervisedfeatureselection;mutualinformation;minimumredundancyandmaximumrelevance;unsupervisedminimumredundancyandmaximumrelevance摘要在数据分析中,特征选择可以用来降低特征的冗余,提高分析结果的可理解性和发现高维数据中隐藏的结构.提出了一种基于互信息的无监督的特征选择方法(UFS-MI),在UFS-MI中,使用了一种综合考虑了相关度和冗余度的特征选择标准UmRMR(无监督最小冗余最大相
8、关)来评价特征的重要性.
此文档下载收益归作者所有