A review of robust clustering

《A review of robust clustering》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、AdvDataAnalClassif(2010)4:89109DOI10.1007/s11634-010-0064-5REGULARARTICLEAreviewofrobustclusteringmethodsLuisAngelGarcía-Escudero·AlfonsoGordaliza·CarlosMatrán·AgustínMayo-IscarReceived:16February2010/Revised:21April2010/Accepted:28June2010/Publishedonline:18June2010©Springer-Verlag2010AbstractDev

2、iationsfromtheoreticalassumptionstogetherwiththepresenceofcertainamountofoutlyingobservationsarecommoninmanypracticalstatisticalappli-cations.ThisisalsothecasewhenapplyingClusterAnalysismethods,wherethosetroublescouldleadtounsatisfactoryclusteringresults.RobustClusteringmethodsareaimedatavoidingth

3、eseunsatisfactoryresults.Moreover,thereexistcertainconnec-tionsbetweenrobustproceduresandClusterAnalysisthatmakeRobustClusteringanappealingunifyingframework.Areviewofdifferentrobustclusteringapproachesintheliteratureispresented.Specialattentionispaidtomethodsbasedontrimmingwhichtrytodiscardmostout

4、lyingdatawhencarryingouttheclusteringprocess.KeywordsClustering·Robustness·Model-basedclustering·TrimmingMathematicsSubjectClassification(2000)62h30·62G35L.A.García-Escudero(B)·A.Gordaliza·C.Matrán·A.Mayo-IscarDepartamentodeEstadísticaeInvestigaciónOperativa,FacultaddeCiencias,UniversidaddeValladol

5、id,PradodelaMagdalena,47002Valladolid,Spaine-mail:lagarcia@eio.uva.esA.Gordalizae-mail:alfonsog@eio.uva.esC.Matráne-mail:matran@eio.uva.esA.Mayo-Iscare-mail:agustinm@eio.uva.es12390L.A.García-Escuderoetal.1IntroductionUnexpecteddeviationsfromtheassumedhypothesestogetherwiththepresenceofcertainamou

6、ntofoutlyingobservationsarecommontroublesinmostpracticalstatis-ticalapplications.ThisisalsothecasewhenusingClusterAnalysistechniques.ClusterAnalysismethodsaimtodetecthomogeneousgroupswithlargeheterogeneityamongthem(see,e.g.,Everitt1977;Gordon1981;KaufmanandRousseeuw1990),butthesearchofgroupsorclus

7、terscanbecompletelyspoiledduetothelackofrobustnessofstandardclusteringmethods.Forinstance,severalveryheterogeneousgroupscanappearartificiallyjoinedtogetherasconstitutingasinglecluster,orsomespuriousclustersmadeupm