文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > A MinMaxCut Spectral Method for Data Clustering and Graph Partitioning
侵权投诉

A MinMaxCut Spectral Method for Data Clustering and Graph Partitioning

A MinMaxCut Spectral Method for Data Clustering and Graph Partitioning

graphpartitioningmethodsRatioCutandNormalizedCutattempttoensurethatobjectsindi erentclusteraremaximallydi erent,whichis(P2).

InthispaperweintroduceMinMaxCut,agraphpartitioningbasedclusteringalgorithmwhichincorporate(P1)and(P2)simultaneously.Weformallystatethemasthefollowingmin-maxclusteringprinciple:datashouldbegroupedintoclusterssuchthatthesimilarityorassociationacrossclustersisminimized,whilethesimilarityorassociationwithineachclusterismaximized(see[14,35]forrecentstudiesofclusteringobjectivefunctions).ClusteringalgorithmssuchasK-meansandthosebasedonGaussianmixturesrequirethecoordinates/attributesofeachobjectexplicitly.Graphpartitioningalgorithmrequiresonlythepairwisesimilaritiesbetweenobjects.GiventhepairwisesimilarityS=(sij),wheresijindicatesthesimilaritybetweenobjectsiandj,wemayconsiderSastheadjacencymatrixofaweightedgraphG,hencethedataclusteringproblembecomesagraphpartitionproblem.(Splittingadatasetintotwoisrephrasedascuttingagraphintotwosubgraphs;cuttingagraphintotwoveryunbalancedsubgraphsisreferredtoasskewedcut;theboundarybetweentwosubgraphsissometimescalledcut.)

Clusteranalysisisappliedtolargeamountofdatawithavarietyofdistributions/shapesfortheclusters.Usingthesimilaritymetric,morecomplexshapeddistributionscanbeaccommodated.ForexampleK-meansfavorssphericallyshapedclusters,whilehierarchicalagglomerativeclusteringcanproduceelongate-shapedclustersbyusingsingle-linkage.Usingsimilarity-basedgraphpartitioning,theconnectivitybetweenobjectsbecomesmostimportant,insteadoftheirshapeinaEuclideanspacewhichisveryhardtomodel.

Themin-maxclusteringprinciplefavorstheobjectivefunctionoptimizationapproach,i.e,clustersareobtainedbyoptimizinganappropriateobjectivefunction.Thisismathematicallymoreprincipledapproach,incontrasttoprocedure-orientedclusteringmethods,suchasthehierarchicalalgorithms.

Inthefollowingwebrie ysummarizetheresultsobtainedinthispaper,whichalsoserveastheoutlineofthepaper.In§2,wediscusstheMinMaxCutforK=2case.We rstshowthatthecontinuoussolutionoftheclustermembershipindicatorvectoristheeigenvectorofthegeneralizedLaplacianmatrixofthesimilaritymatrix.Relatedworkonspectralgraphclustering,theRatioCut[19]andtheNormalizedCut[32]arediscussedin§2.1.Usingrandomgraphmodel,weshowthatMinMaxCuttendstoproducebalancedclusterswhileearliermethodsdonot(see§2.2).TheclusterbalancingpowerofMinMaxCutcanbesoftenedorhardenedbyaslightgeneralizationoftheclusteringobjectivefunction(see§2.3).In§2.4,wede nethecohesionofadataset/graphastheoptimalvalueofMinMaxCutobjectivefunctionwhenthedatasetissplitintotwo.Weproveimportantlowerandupperboundsforthecohesionvalue.Experimentsonclusteringinternetnewsgroupsarepresentedin§2.5.whichshowtheadvantageofMinMaxCutcomparedwithexistingmethods.In§2.6wederivetheconditionsforpossibleskewedclusteringforMinMaxCutandNormalizedCutwhichshowsthebalancingpowerofMinMaxCut.In§2.7,weshowtheMinMaxCutlinkageisusefulforfurtherre nementofclustersobtainedfromMinMaxCut.Thelinkagedi erentialorderingcanfurtherimprovetheclusteringresults(see§2.8).In§2.9,wediscusstheclusteringofacontingencytablewhichcanbeviewedasaweightedbipartitegraph.Thesimultaneousclusteringofrowsandcolumnsofthecontingencytablecanbedoneinmuchthesamewayasin2-wayclusteringof§2.In§3,wediscussMinMaxCutforK>2cases.WeshowK-wayMinMaxCutleadstothemorere nedorsubtleclusterbalance,thesimilarity-weightedsizebalancein§3.1.In§3.2,theimportanceofKeigenvectorsarenotedwiththegeneralizedlowerandupperboundsofoptimalvalueoftheobjectivefunction.TheK-wayclusteringrequirestwostages,theinitialclusteringandre nement.In§3.3,threemethodsofinitialclusteringarebrie yexplained,theeigenspaceK-means,thedivisiveandagglomerativeclusterings.Theclusterre nementalgorithmsbasedonMinMaxCutobjectivefunctionsareoutlinein§3.4.

上一页第2页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email