Bayesian Statistics

  • 格式:pdf
  • 大小:77.89 KB
  • 文档页数:5

HeckermanD1996Bayesiannetworksfordatamining.DataMiningandKnowledgeDisco󰀁ery1:79–119HeckermanD,GeigerD1995LearningBayesiannetworks:AunificationfordiscreteandGaussiandomains.In:Proc.11thConf.onUncertaintyinArtificialIntelligence.MorganKauf-mann,SanFrancisco,pp.274–84.SeealsoTechnicalReportTR-95-16,MicrosoftResearch,Redmond,WA,February1995HowardR,MathesonJ1981Influencediagrams.In:HowardR,MathesonJ(eds.)ReadingsonthePrinciplesandApplicationsofDecisionAnalysis.StrategicDecisionsGroup,MenloPark,CA,Vol.II,pp.721–62JordanM(ed.)1998LearninginGraphicalModels.Kluwer,DordrechtLauritzenS1992Propagationofprobabilities,means,andvariancesinmixedgraphicalassociationmodels.JournaloftheAmericanStatisticalAssociation87:1098–108LauritzenS1996GraphicalModels.ClarendonPress,Oxford,UKLauritzenS,SpiegelhalterD1988Localcomputationswithprobabilitiesongraphicalstructuresandtheirapplicationtoexpertsystems.JournaloftheRoyalStatisticalSocietyB50:157–224PearlJ1988ProbabilisticReasoninginIntelligentSystems:NetworksofPlausibleInference.MorganKaufmann,SanMateo,CAPearlJ(ed.)2000Causality:Models,Reasoning,andInference.CambridgeUniversityPress,Cambridge,UKShachterR1988Probabilisticinferenceandinfluencediagrams.OperationsResearch36:589–604SpiegelhalterD,ThomasA1998Graphicalmodelingforcomplexstochasticsystems:TheBUGSproject.IEEEIn-telligentSystemsandtheirApplications13:14–5SpirtesP,GlymourC,ScheinesR2001Causation,Prediction,andSearch,2ndedn.MITPress,Cambridge,MAWermuthN1976Analogiesbetweenmultiplicativemodelsincontingencytablesandcovarianceselection.Biometrics32:95–108WhittakerJ1990GraphicalModelsinAppliedMulti󰀁ariateStatistics.Wiley,NewYorkWrightS1921Correlationandcausation.JournalofAgriculturalResearch20:557–85

D.Heckerman

BayesianStatistics

Bayesianstatisticsreferstoanapproachtostatistical

inferencecharacterizedbytwokeyideas:(a)all

unknownquantities,includingparameters,aretreated

asrandomvariableswithprobabilitydistributions

usedtodescribethestateofknowledgeaboutthe

valuesoftheseunknowns,and(b)statisticalinferences

abouttheunknownquantitiesbasedonobserveddata

arederivedusingBayes’theorem(describedbelow).

TheBayesianapproachsharesmanyfeatureswiththe

traditionalfrequentistapproachtoinference(e.g.,the

useofparametricmodels,thatistosay,models

dependentonunknownparameters,fordescribing

data)butdiffersinitsrelianceonprobabilitydistri-

butionsforunknowns(includingtheparameters).Thetraditionalapproachtoinferencereliesontherepeated

samplingdistributionofthedataforafixedbut

unknownparametervalue,essentiallyaskingwhat

wouldhappenifmanynewsamplesweredrawn;the

Bayesianapproachtreatstheparameterasarandom

variableandassignsitaprobabilitydistribution.

Qualitatively,theBayesianapproachtoinference

beginswithaprobabilitydistributiondescribingthe

stateofknowledgeaboutunknownquantities(usually

parameters)beforecollectingdata,andthenuses

observeddatatoupdatethisdistribution.Inthis

articlethebasicelementsofaBayesiananalysisare

reviewed:modelspecification,calculationofthepos-

teriordistribution,modelchecking,andsensitivity

analysis.Additionalsectionsaddressthechoiceof

priordistribution,andtheapplicationofBayesian

methods.Additionaldetailsaboutmostofthetopics

inthisarticlecanbefoundinthebooksbyO’Hagan

(1994),Gelman(1995),CarlinandLouis(2000),

andGilksetal.(1998).

Theearliestdevelopmentsrelatedtotheapplication

ofprobabilitytoquestionsofinferencedatetothe

contributionsofBayesandLaplaceinthesecondhalf

oftheeighteenthcentury(Stigler1986).Priortothat

pointresearchersfocusedonthetraditionalpre-data

probabilitycalculations,i.e.,givencertainassump-

tionsabouttherandomprocess,whatistheprob-

abilityassignedtovariouspossibleoutcomesfora

variableinquestion?BayesandLaplacereceive

independentcreditfor‘inverting’theprobability

statementtomakeprobabilitystatementsaboutpar-

ametervalues,givenobserveddatavalues.Therewas

littleactivityafterthattime,thoughsomeindividuals,

notablythephysicistJeffreys(1961),continuedto

developthefieldofinductiveinference.Modern

Bayesianinferencedevelopedintheperiodaround

andafterWorldWarII(e.g.,seeStatistics:TheField).

ThenameBayesianinferencereplaces‘inverseprob-

ability’onlyatthislatertime.Somekeycontributions:

Savage(1954)isaninfluentialbookusingdecision

theorytojustifyBayesianmethods;deFinetti(1974)

contributedcrucialworkconcerningtheroleof

exchangeability(whichplaysaroleanalogoustothat

ofindependentidenticallydistributedobservationsin

thetraditionalfrequentistapproachtoinference);

RaiffaandSchlaifer(1961)developedtheuseof

conjugatedistributionsindetail;Lindley(1971,1990)

andBoxandTiao(1973)contributedgreatlytothe

popularizationoftheapproach.Mostrecently,thelast

decadeofthetwentiethcenturysawthediscovery(or

rediscovery)ofcomputationalalgorithmsthatmakeit