聚类分析外文文献及翻译.docx

聚类分析外文文献及翻译.docx

《聚类分析外文文献及翻译.docx》由会员分享，可在线阅读，更多相关《聚类分析外文文献及翻译.docx（15页珍藏版）》请在冰豆网上搜索。

聚类分析外文文献及翻译

聚类分析外文文献及翻译

本科毕业论文

外文文献及译文

文献、资料题目：

ClusterAnalysis

—BasicConceptsandAlgorithms

文献、资料来源：

文献、资料发表（出版）日期：

院（部）：

土木工程学院

专业：

土木工程

班级：

姓名：

学号：

指导教师：

翻译日期：

外文文献：

ClusterAnalysis

—BasicConceptsandAlgorithms

Clusteranalysisdividesdataintogroups（clusters）thataremeaningful,useful,orboth.Ifmeaningfulgroupsarethegoal,thentheclustersshouldcapturethenaturalstructureofthedata.Insomecases,however,clusteranalysisisonlyausefulstartingpointforotherpurposes,suchasdatasummarization.Whetherforunderstandingorutility,clusteranalysishaslongplayedanimportantroleinawidevarietyoffields:

psychologyandothersocialsciences,biology,statistics,patternrecognition,informationretrieval,machinelearning,anddatamining.

Therehavebeenmanyapplicationsofclusteranalysistopracticalproblems.Weprovidesomespecificexamples,organizedbywhetherthepurposeoftheclusteringisunderstandingorutility.

ClusteringforUnderstandingClasses,orconceptuallymeaningfulgroupsofobjectsthatsharecommoncharacteristics,playanimportantroleinhowpeopleanalyzeanddescribetheworld.Indeed,humanbeingsareskilledatdividingobjectsintogroups（clustering）andassigningparticularobjectstothesegroups（classification）.Forexample,evenrelativelyyoungchildrencanquicklylabeltheobjectsinaphotographasbuildings,vehicles,people,animals,plants,etc.Inthecontextofunderstandingdata,clustersarepotentialclassesandclusteranalysisisthestudyoftechniquesforautomaticallyfindingclasses.Thefollowingaresomeexamples:

Biology.Biologistshavespentmanyyearscreatingataxonomy（hierarchicalclassification）ofalllivingthings:

kingdom,phylum,class,order,family,genus,andspecies.Thus,itisperhapsnotsurprisingthatmuchoftheearlyworkinclusteranalysissoughttocreateadisciplineofmathematicaltaxonomythatcouldautomaticallyfindsuchclassificationstructures.Morerecently,biologistshaveappliedclusteringtoanalyzethelargeamountsofgeneticinformationthatarenowavailable.Forexample,clusteringhasbeenusedtofindgroupsofgenesthathavesimilarfunctions.

•InformationRetrieval.TheWorldWideWebconsistsofbillionsofWebpages,andtheresultsofaquerytoasearchenginecanreturnthousandsofpages.Clusteringcanbeusedtogroupthesesearchresultsintoasmallnumberofclusters,eachofwhichcapturesaparticularaspectofthequery.Forinstance,aqueryof“movie”mightreturnWebpagesgroupedintocategoriessuchasreviews,trailers,stars,andtheaters.Eachcategory（cluster）canbebrokenintosubcategories（sub-clusters）,producingahierarchicalstructurethatfurtherassistsauser’sexplorationofthequeryresults.

•Climate.UnderstandingtheEarth’sclimaterequiresfindingpatternsintheatmosphereandocean.Tothatend,clusteranalysishasbeenappliedtofindpatternsintheatmosphericpressureofpolarregionsandareasoftheoceanthathaveasignificantimpactonlandclimate.

•PsychologyandMedicine.Anillnessorconditionfrequentlyhasanumberofvariations,andclusteranalysiscanbeusedtoidentifythesedifferentsubcategories.Forexample,clusteringhasbeenusedtoidentifydifferenttypesofdepression.Clusteranalysiscanalsobeusedtodetectpatternsinthespatialortemporaldistributionofadisease.

•Business.Businessescollectlargeamountsofinformationoncurrentandpotentialcustomers.Clusteringcanbeusedtosegmentcustomersintoasmallnumberofgroupsforadditionalanalysisandmarketingactivities.

ClusteringforUtility：

Clusteranalysisprovidesanabstractionfromindividualdataobjectstotheclustersinwhichthosedataobjectsreside.Additionally,someclusteringtechniquescharacterizeeachclusterintermsofaclusterprototype;i.e.,adataobjectthatisrepresentativeoftheotherobjectsinthecluster.Theseclusterprototypescanbeusedasthebasisforanumberofdataanalysisordataprocessingtechniques.Therefore,inthecontextofutility,clusteranalysisisthestudyoftechniquesforfindingthemostrepresentativeclusterprototypes.

•Summarization.Manydataanalysistechniques,suchasregressionorPCA,haveatimeorspacecomplexityofO（m2）orhigher（wheremisthenumberofobjects）,andthus,arenotpracticalforlargedatasets.However,insteadofapplyingthealgorithmtotheentiredataset,itcanbeappliedtoareduceddatasetconsistingonlyofclusterprototypes.Dependingonthetypeofanalysis,thenumberofprototypes,andtheaccuracywithwhichtheprototypesrepresentthedata,theresultscanbecomparabletothosethatwouldhavebeenobtainedifallthedatacouldhavebeenused.

•Compression.Clusterprototypescanalsobeusedfordatacompres-sion.Inparticular,atableiscreatedthatconsistsoftheprototypesforeachcluster;i.e.,eachprototypeisassignedanintegervaluethatisitsposition（index）inthetable.Eachobjectisrepresentedbytheindexoftheprototypeassociatedwithitscluster.Thistypeofcompressionisknownasvectorquantizationandisoftenappliedtoimage,sound,andvideodata,where

（1）manyofthedataobjectsarehighlysimilartooneanother,

（2）somelossofinformationisacceptable,and（3）asubstantialreductioninthedatasizeisdesired

•EffcientlyFindingNearestNeighbors.Findingnearestneighborscanrequirecomputingthepairwisedistancebetweenallpoints.Oftenclustersandtheirclusterprototypescanbefoundmuchmoreeffciently.Ifobjectsarerelativelyclosetotheprototypeoftheircluster,thenwecanusetheprototypestoreducethenumberofdistancecomputationsthatarenecessarytofindthenearestneighborsofanobject.Intuitively,iftwoclusterprototypesarefarapart,thentheobjectsinthecorrespondingclusterscannotbenearestneighborsofeachother.Consequently,tofindanobject’snearestneighborsitisonlynecessarytocomputethedistancetoobjectsinnearbyclusters,wherethenearnessoftwoclustersismeasuredbythedistancebetweentheirprototypes.

Thischapterprovidesanintroductiontoclusteranalysis.Webeginwithahigh-leveloverviewofclustering,includingadiscussionofthevariousap-proachestodividingobjectsintosetsofclustersandthedifferenttypesofclusters.Wethendescribethreespecificclusteringtechniquesthatrepresentbroadcategoriesofalgorithmsandillustrateavarietyofconcepts:

K-means,agglomerativehierarchicalclustering,andDBSCAN.Thefinalsectionofthischapterisdevotedtoclustervalidity—methodsforevaluatingthegoodnessoftheclustersproducedbyaclusteringalgorithm.MoreadvancedclusteringconceptsandalgorithmswillbediscussedinChapter9.Wheneverpossible,wediscussthestrengthsandweaknessesofdifferentschemes.Inaddition,thebibliographicnotesprovidereferencestorelevantbooksandpapersthatexploreclusteranalysisingreaterdepth.

1.1Overview

Beforediscussingspecificclusteringtechniques,weprovidesomenecessarybackground.First,wefurtherdefineclusteranalysis,illustratingwhyitisdiffcultandexplainingitsrelationshiptoothertechniquesthatgroupdata.Thenweexploretwoimportanttopics:

（1）differentwaystogroupasetofobjectsintoasetofclusters,and

（2）typesofclusters.

1.1.1WhatIsClusterAnalysis?

Clusteranalysisgroupsdataobjectsbasedonlyoninformationfoundinthedatathatdescribestheobjectsandtheirrelationships.Thegoalisthattheobjectswithinagroupbesimilar（orrelated）tooneanotheranddifferentfrom（orunrelatedto）theobjectsinothergroups.Thegreaterthesimilarity（orhomogeneity）withinagroupandthegreaterthedifferencebetweengroups,thebetterormoredistincttheclustering.

Clusteranalysisisrelatedtoothertechniquesthatareusedtodividedataobjectsintogroups.Forinstance,clusteringcanberegardedasaformofclassificationinthatitcreatesalabelingofobjectswithclass（cluster）labels.However,itderivestheselabelsonlyfromthedata.Incontrast,classificationnthesenseofChapter4issupervisedclassification;i.e.,new,unlabeledobjectsareassignedaclasslabelusingamodeldevelopedfromobjectswithknownclasslabels.Forthisreason,clusteranalysisissometimesreferredtoasunsupervisedclassification.Whenthetermclassificationisusedwithoutanyqualificationwithindatamining,ittypicallyreferstosupervisedclassification.

Also,whilethetermssegmentationandpartitioningaresometimesusedassynonymsforclustering,thesetermsarefrequentlyusedforapproachesoutsidethetraditionalboundsofclusteranalysis.Forexample,thetermpartitioningisoftenusedinconnectionwithtechniquesthatdividegraphsintosubgraphsandthatarenotstronglyconnectedtoclustering.Segmentationoftenreferstothedivisionofdataintogroupsusingsimpletechniques;e.g.,animagecanbesplitintosegmentsbasedonlyonpixelintensityandcolor,orpeoplecanbedividedintogroupsbasedontheirincome.Nonetheless,someworkingraphpartitioningandinimageandmarketsegmentationisrelatedtoclusteranalysis.

1.1.2DifferentTypesofClusterings

Anentirecollectionofclustersiscommonlyreferredtoasaclustering,andinthissection,wedistinguishvarioustypesofclusterings:

hierarchical（nested）versuspartitional（unnested）,exclusiveversusoverlappingversusfuzzy,andcompleteversuspartial.

HierarchicalversusPartitionalThemostcommonlydiscusseddistinc-tionamongdifferenttypesofclusteringsiswhetherthesetofclustersisnestedorunnested,orinmoretraditionalterminology,hierarchicalorpartitional.Apartitionalclusteringissimply