[5] Statistical Modeling The Two Cultures Leo Breiman.pdf

《[5] Statistical Modeling The Two Cultures Leo Breiman.pdf》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、StatisticalScience2001,Vol.16,No.3,199–231StatisticalModeling:TheTwoCulturesLeoBreimanAbstract.Therearetwoculturesintheuseofstatisticalmodelingtoreachconclusionsfromdata.Oneassumesthatthedataaregeneratedbyagivenstochasticdatamodel.Theotherusesalgorithmicmode

2、lsandtreatsthedatamechanismasunknown.Thestatisticalcommunityhasbeencommittedtothealmostexclusiveuseofdatamodels.Thiscommit-menthasledtoirrelevanttheory,questionableconclusions,andhaskeptstatisticiansfromworkingonalargerangeofinterestingcurrentprob-lems.Algor

3、ithmicmodeling,bothintheoryandpractice,hasdevelopedrapidlyinfieldsoutsidestatistics.Itcanbeusedbothonlargecomplexdatasetsandasamoreaccurateandinformativealternativetodatamodelingonsmallerdatasets.Ifourgoalasafieldistousedatatosolveproblems,thenweneedtomoveaway

4、fromexclusivedependenceondatamodelsandadoptamorediversesetoftools.1.INTRODUCTIONThevaluesoftheparametersareestimatedfromthedataandthemodelthenusedforinformationStatisticsstartswithdata.Thinkofthedataasand/orprediction.Thustheblackboxisfilledinlikebeinggenerat

5、edbyablackboxinwhichavectorofthis:inputvariablesx(independentvariables)goinoneside,andontheothersidetheresponsevariablesylinearregressionyxcomeout.Insidetheblackbox,naturefunctionstologisticregressionassociatethepredictorvariableswiththeresponseCoxmodelvaria

6、bles,sothepictureislikethis:Modelvalidation.Yes–nousinggoodness-of-fittestsandresidualexamination.ynaturexEstimatedculturepopulation.98%ofallstatisti-cians.Therearetwogoalsinanalyzingthedata:TheAlgorithmicModelingCulturePrediction.Tobeabletopredictwhattheresp

7、onsesaregoingtobetofutureinputvariables;TheanalysisinthiscultureconsiderstheinsideofInformation.Toextractsomeinformationabouttheboxcomplexandunknown.Theirapproachistohownatureisassociatingtheresponsevariablesfindafunctionfx—analgorithmthatoperatesontotheinp

8、utvariables.xtopredicttheresponsesy.Theirblackboxlookslikethis:Therearetwodifferentapproachestowardthesegoals:yunknownxTheDataModelingCultureTheanalysisinthisculturestartswithassumingdecisiontre