CHAMELEON A Hierarchical Clustering Algorithm ：变色龙的层次聚类算法ppt课件.ppt

《CHAMELEON A Hierarchical Clustering Algorithm ：变色龙的层次聚类算法ppt课件.ppt》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、CHAMELEON:AHierarchicalClusteringAlgorithmUsingDynamicModelingPaperpresentationindataminingclassPresenter:許明壽;蘇建仲Data:2001/12/182001/12/181CHAMELEONAboutthispaper…DepartmentofComputerScienceandEngineering,UniversityofMinnesotaGeorgeKarypisEui-Honh(Sam)HanVipinKumarIEEEComputerJour

2、nal-Aug.1999OutlineProblemsdefinitionMainalgorithmKeysfeaturesofCHAMELEONExperimentandrelatedworkedConclusionanddiscussionProblemsdefinitionClusteringIntraclustersimilarityismaximizedInterclustersimilarityisminimizedProblemsofexistingclusteringalgorithmsStaticmodelconstrainBreakdo

3、wnwhenclustersthatareofdiverseshapes，densities，andsizesSusceptibletonoise,outliers,andartifactsStaticmodelconstrainDataspaceconstrainKmeans,PAM…etcSuitableonlyfordatainmetricspacesClustershapeconstrainKmeans,PAM,CLARANSAssumeclusterasellipsoidalorglobularandaresimilarsizesClusterd

4、ensityconstrainDBScanPointswithingenuineclusteraredensity-reachableandpointacrossdifferentclustersarenotSimilaritydetermineconstrainCURE,ROCKUsestaticmodeltodeterminethemostsimilarclustertomergePartitiontechniquesproblem(a)Clustersofwidelydifferentsizes(b)ClusterswithconvexshapesH

5、ierarchicaltechniqueproblem(1/2)The{(c),(d)}willbechoosetomergewhenweonlyconsiderclosenessHierarchicaltechniqueproblem(2/2)The{(a),(c)}willbechoosetomergewhenweonlyconsiderinter-connectivityMainalgorithmTwophasealgorithmPHASEIUsegraphpartitioningalgorithmtoclusterthedataitemsintoa

6、largenumberofrelativelysmallsub-clusters.PHASEIIUsesanagglomerativehierarchicalclusteringalgorithmtofindthegenuineclustersbyrepeatedlycombiningtogetherthesesub-clusters.FrameworkConstructSparseGraphPartitiontheGraphMergePartitionFinalClustersDataSetKeysfeaturesofCHAMELEONModelingt

7、hedataModelingtheclustersimilarityPartitionalgorithmsMergeschemesTermsArgumentsneededKK-nearestneighborgraphMINSIZETheminimasizeofinitialclusterTRIThresholdofrelatedinter-connectivityTRCThresholdofrelatedintra-connectivityαCoefficientforweightofRIandRCModelingthedataK-nearestneigh

8、borgraphapproachAdvantagesDatapoi