【OReilly】如何利用合成数据推进人工智能和机器学习_第1页
【OReilly】如何利用合成数据推进人工智能和机器学习_第2页
【OReilly】如何利用合成数据推进人工智能和机器学习_第3页
【OReilly】如何利用合成数据推进人工智能和机器学习_第4页
【OReilly】如何利用合成数据推进人工智能和机器学习_第5页
已阅读5页,还剩106页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

Complimentsof

Accelerating

AIwith

SyntheticData

GeneratingDataforAIProjects

KhaledElEmam

THELEADERINAICOMPUTING.

Signuptogetthe

latestAInewsstraight

toyourinbox.

SUBSCRIBE

AcceleratingAIwith SyntheticDataGeneratingDataforAIProjects

KhaledElEmam

Beijing·Boston·Farnham·Sebastopol·Tokyo

AcceleratingAIwithSyntheticData

byKhaledElEmam

Copyright©2020O’ReillyMedia,Inc.Allrightsreserved.

PrintedintheUnitedStatesofAmerica.

PublishedbyO’ReillyMedia,Inc.,1005GravensteinHighwayNorth,Sebastopol,CA95472.

O’Reillybooksmaybepurchasedforeducational,business,orsalespromotionaluse.Onlineeditionsarealsoavailableformosttitles

(

).Formoreinfor‐mation,contactourcorporate/institutionalsalesdepartment:800-998-9938or

corporate@

.

AcquisitionsEditor:JonathanHassell

DevelopmentEditor:MelissaPotter

ProductionEditor:DanielElfanbaum

Copyeditor:SharonWilkey

Proofreader:ShannonTurlington

InteriorDesigner:DavidFutato

CoverDesigner:KarenMontgomery

Illustrator:RebeccaDemarest

June2020:FirstEdition

RevisionHistoryfortheFirstEdition

2020-06-03:FirstRelease

TheO’ReillylogoisaregisteredtrademarkofO’ReillyMedia,Inc.AcceleratingAIwithSyntheticData,thecoverimage,andrelatedtradedressaretrademarksofO’ReillyMedia,Inc.

Theviewsexpressedinthisworkarethoseoftheauthor,anddonotrepresentthepublisher’sviews.Whilethepublisherandtheauthorhaveusedgoodfaitheffortstoensurethattheinformationandinstructionscontainedinthisworkareaccurate,thepublisherandtheauthordisclaimallresponsibilityforerrorsoromissions,includ‐ingwithoutlimitationresponsibilityfordamagesresultingfromtheuseoforreli‐anceonthiswork.Useoftheinformationandinstructionscontainedinthisworkisatyourownrisk.Ifanycodesamplesorothertechnologythisworkcontainsordescribesissubjecttoopensourcelicensesortheintellectualpropertyrightsofoth‐ers,itisyourresponsibilitytoensurethatyourusethereofcomplieswithsuchlicen‐sesand/orrights.

ThisworkispartofacollaborationbetweenO’ReillyandNVIDIA.Seeour

state‐

mentofeditorialindependence

.

978-1-492-04596-0

[LSI]

v

TableofContents

1.

DefiningSyntheticData 1

WhatIsSyntheticData?

2

TheBenefitsofSyntheticData

5

LearningtoTrustSyntheticData

9

OtherApproachestoAccessingData

11

GeneratingSyntheticDatafromRealData

12

Conclusions

15

2.

TheSynthesisProcess 17

DataSynthesisProjects

17

TheDataSynthesisPipeline

21

SynthesisProgramManagement

27

BestPracticesforImplementingDataSynthesis

28

Conclusions

30

3.

SyntheticDataCaseStudies 33

ManufacturingandDistribution

34

HealthCare

36

FinancialServices

43

Transportation

46

Conclusions

50

4.

TheFutureofDataSynthesis 51

CreatingaDataUtilityFramework

51

RemovingInformationfromSyntheticData

52

vi|TableofContents

UsingDataWatermarking

53

GeneratingSynthesisfromSimulators

54

Conclusions

55

CHAPTER1

1

DefiningSyntheticData

Interestinsyntheticdatahasbeengrowingquiterapidlyoverthelastfewyears.Thishasbeendrivenbytwosimultaneoustrends.Thefirstisthedemandforlargeamountsofdatatotrainandbuildarti-ficialintelligenceandmachinelearning(AIML)models.Thesecondisrecentworkthathasdemonstratedeffectivemethodstogeneratehigh-qualitysyntheticdata.Bothhaveresultedintherecognitionthatsyntheticdatacansolvesomedifficultproblemsquiteeffec-tively,especiallywithintheAIMLcommunity.Groupsandbusi-nesseswithincompanieslikeNVIDIA,IBM,andAlphabet,aswellasagenciessuchastheUSCensusBureau,haveadopteddifferenttypesofdatasynthesistosupportmodelbuilding,applicationdevel-opment,anddatadissemination.

Thisreportprovidesageneraloverviewofsyntheticdatageneration,withafocusonthebusinessvalueandusecases,andhigh-levelcov-erageoftechniquesandimplementationpractices.Weaimtoanswerthequestionsthatabusinessreaderwouldtypicallyask(andhastypicallyasked),butatthesametimeprovidesomedirectiontoanalyticsleadershipseekingtounderstandtheoptionsavailableandwheretolooktogetstarted.

WeshowhowsyntheticdatacanaccelerateAIMLprojects.Someproblemsthatcanbetackledbyusingsyntheticdatawouldbetoocostlyordangerous(e.g.,inthecaseoftrainingmodelscontrollingautonomousvehicles)tosolveusingmoretraditionalmethods,orsimplycannotbedoneotherwise.

2|Chapter1:DefiningSyntheticData

AIMLprojectsrunindifferentindustries,andthemultipleindustryusecasesthatweincludeinthisreportareintendedtogiveyouaflavorofthebroadapplicationsofdatasynthesis.WedefineanAIMLprojectquitebroadlyaswell,toinclude,forexample,thedevelopmentofsoftwareapplicationsthathaveAIMLcomponents.Thereportisdividedintofourchapters.Thisintroductorychaptercoversbasicconceptsandpresentsthecaseforsyntheticdata.

Chap‐

ter2

presentsthedatasynthesisprocessandpipelines,scalingimplementationintheenterprise,andbestpractices.Aseriesofindustry-specificcasestudiesfollowin

Chapter3

.

Chapter4

isforward-lookingandconsiderswherethistechnologyisheaded.

Inthischapter,westartbydefiningthetypesofsyntheticdata.Thisisfollowedbyadescriptionofthebenefitsofusingsyntheticdata—thetypesofproblemsthatdatasynthesiscansolve.Giventherecentadoptionofthisapproachintopractice,buildingtrustinanalysisresultsfromsyntheticdataisimportant.Wethereforealsopresentexamplessupportingtheutilityofsyntheticdataanddiscussmeth‐odstobuildtrust.

Alternativestodatasynthesisexist,andwepresentthesenextwithanassessmentofstrengthsandweaknesses.Thischapterthencloseswithanoverviewofmethodsforsyntheticdatageneration.

WhatIsSyntheticData?

Ataconceptuallevel,syntheticdataisnotrealdatabutisdatathathasbeengeneratedfromrealdataandthathasthesamestatisticalpropertiesastherealdata.Thismeansthatananalystwhoworkswithasyntheticdatasetshouldgetanalysisresultsthataresimilartothosetheywouldgetwithrealdata.Thedegreetowhichasyn‐theticdatasetisanaccurateproxyforrealdataisameasureofutil-ity.Furthermore,werefertotheprocessofgeneratingsyntheticdataassynthesis.

Datainthiscontextcanmeandifferentthings.Forexample,datacanbestructureddata(i.e.,rowsandcolumns),asonewouldseeinarelationaldatabase.Datacanalsobeunstructuredtext,suchasdoc‐tors’notes,transcriptsofconversationsamongpeopleorwithdigitalassistants,oronlineinteractionsbyemailorchat.Furthermore,images,videos,audio,andvirtualenvironmentsarealsotypesofdatathatcanbesynthesized.Wehaveseenexamplesoffakeimages

WhatIsSyntheticData?|3

inthemachinelearningliterature;forinstance,realisticfacesofpeoplewhodonotexistintherealworldcanbecreated,andyoucan

viewtheresults

online.

Syntheticdataisdividedintotwotypes,basedonwhetheritisgen‐eratedfromactualdatasetsornot.

Thefirsttypeissynthesizedfromrealdatasets.Theanalystwillhavesomerealdatasetsandthenbuildamodeltocapturethedistribu‐tionsandstructureofthatrealdata.Here,structuremeansthemul‐tivariaterelationshipsandinteractionsinthedata.Thenthesyntheticdataissampledorgeneratedfromthatmodel.Ifthemodelisagoodrepresentationoftherealdata,thesyntheticdatawillhavesimilarstatisticalpropertiesastherealdata.

Forexample,adatasciencegroupspecializinginunderstandingcustomerbehaviorswouldneedlargeamountsofdatatobuilditsmodels.Butbecauseofprivacyorotherconcerns,theprocessforgettingaccesstothatcustomerdataisslowanddoesnotprovidegoodenoughdatawhenitdoesarrivebecauseofextensivemaskingandredactionofinformation.Instead,asyntheticversionoftheproductiondatasetscanbeprovidedtotheanalystsforbuildingtheirmodels.Thesynthesizeddatawillhavefewerconstraintsputonitsuseandwouldallowthemtoprogressmorerapidly.

Thesecondtypeofsyntheticdataisnotgeneratedfromrealdata.Itiscreatedbyusingexistingmodelsorbyusingbackgroundknowl‐edgeoftheanalyst.Theseexistingmodelscanbestatisticalmodelsofaprocess(forexample,developedthroughsurveysorotherdatacollectionmechanisms)ortheycanbesimulations.Simulationscanbecreated,forinstance,bygamingenginesthatcreatesimulated(andsynthetic)imagesofscenesorobjects,orbysimulationenginesthatgenerateshopperdatawithparticularcharacteristics(say,ageandgender)ofpeoplewhowalkpastthesiteofaprospectivestoreatdifferenttimesoftheday.

Backgroundknowledgecanbe,forexample,amodelofhowafinancialmarketbehavesbasedontextbookdescriptionsorbasedonthebehaviorsofstockpricesundervarioushistoricalconditions,oritcanbeknowledgeofthestatisticaldistributionofhumantrafficinastorebasedonyearsofexperience.Insuchacase,itisrelativelystraightforwardtocreateamodelandsamplefromittogeneratesyntheticdata.Iftheanalyst’sknowledgeoftheprocessisaccurate,thesyntheticdatawillbehaveinamannerthatisconsistentwith

4|Chapter1:DefiningSyntheticData

real-worlddata.Ofcourse,thisworksonlywhenthephenomenonofinterestistrulywellunderstood.

Asafinalexample,whenaprocessisnewornotwellunderstoodbytheanalystandthereisnorealhistoricaldatatouse,ananalystcanmakesomesimpleassumptionsaboutthedistributionsandcorrela-tionsamongthevariablesinvolvedintheprocess.Forexample,theanalystcanmakeasimplifyingassumptionthatthevariableshavenormaldistributionsand“medium”correlationsamongthem,andcreatedatathatway.Thistypeofdatawilllikelynothavethesamepropertiesasrealdatabutcanstillbeusefulforsomepurposes,suchasdebugginganRdataanalysisprogramorforsometypesofper-formancetestingofsoftwareapplications.

Forsomeusecases,havinghighutilitywillmatterquiteabit.Inothercases,mediumorevenlowutilitymaybeacceptable.Forexample,iftheobjectiveistobuildAIMLmodelstopredictcus-tomerbehaviorandmakemarketingdecisionsbasedonthat,highutilitywillbeimportant.Ontheotherhand,iftheobjectiveistoseeifyoursoftwarecanhandlealargevolumeoftransactions,thedatautilityexpectationswillbeconsiderablyless.Therefore,understand-ingwhatdata,models,simulators,andknowledgeexistaswellastherequirementsfordatautilitywilldrivethespecificapproachtouseforgeneratingthesyntheticdata.

Table1-1

providesasummaryofthesyntheticdatatypes.

Table1-1.Typesofdatasynthesiswiththeirutilityandprivacyimplications

Typeofsyntheticdata

Utility

Generatedfromreal(nonpublic)datasetsGeneratedfromrealpublicdata

Canbequitehigh

Canbehigh,althoughlimitationsexistbecause

publicdatatendstobede-identifiedoraggregated

Generatedfromanexistingmodelofa

process,whichcanalsoberepresentedinasimulationengine

Basedonanalystknowledge

Willdependonthefidelityoftheexistinggeneratingmodel

Willdependonhowwelltheanalystknowsthedomainandthecomplexityofthephenomenon

Generatedfromgenericassumptionsnotspecifictothephenomenon

Willlikelybelow

TheBenefitsofSyntheticData|5

Nowthatyouhaveanunderstandingofthetypesofsyntheticdata,wewilllookatthebenefitsofdatasynthesisoverallandforsomeofthesedatatypesspecifically.

TheBenefitsofSyntheticData

Inthissection,wepresentseveralwaysthatdatasynthesiscansolvepracticalproblemswithAIMLprojects.Thebenefitsofsyntheticdatacanbedramatic.Itcanmakeimpossibleprojectsdoable,signif‐icantlyaccelerateAIMLinitiatives,orresultinmaterialimprove‐mentintheoutcomesofAIMLprojects.

ImprovingDataAccess

DataaccessiscriticaltoAIMLprojects.Thedataisneededtotrainandvalidatemodels.Morebroadly,dataisalsoneededforevaluat‐ingAIMLtechnologiesthathavebeendevelopedbyothers,aswellasfortestingAIMLsoftwareapplicationsorapplicationsthatincor‐porateAIMLmodels.

Typically,dataiscollectedforaparticularpurposewiththeconsentoftheindividual;forexample,forparticipatinginawebinarorforparticipatinginaclinicalresearchstudy.Ifyouwanttousethatsamedataforadifferentpurpose,suchasforbuildingamodeltopredictwhatkindofpersonislikelytosignupforawebinarorwhowouldparticipateinastudy,thenthatisconsideredasecondarypurpose.

Accesstodataforsecondaryanalysisisbecomingproblematic.TheUSGovernmentAccountabilityOffice

1

andtheMcKinseyGlobalInstitute

2

bothnotethataccessingdataforbuildingandtestingAIMLmodelsisachallengefortheiradoptionmorebroadly.ADeloitteanalysisconcludedthatdataaccessissuesarerankedinthetopthreechallengesfacedbycompanieswhenimplementingAI.

3

ArecentsurveyfromMITTechnologyReviewreportedthatalmost

1GovernmentAccountabilityOffice,“ArtificialIntelligence:EmergingOpportunities,Challenges,andImplications,”GAO-18-142SP(March2018).

https://oreil.ly/Cpyli

.

2McKinseyGlobalInstitute,“ArtificialIntelligence:TheNextDigitalFrontier?”(June2017).

https://oreil.ly/zJ8oZ

.

3DeloitteInsights,“StateofAIintheEnterprise,2ndEdition”(2018).

https://oreil.ly/

l07tJ

.

6|Chapter1:DefiningSyntheticData

halfoftherespondentsidentifieddataavailabilityasaconstrainttotheuseofAIwiththeircompany.

4

Atthesametime,thepublicisgettinguneasyabouthowtheirdataisusedandshared,andprivacylawsarebecomingmorestrict.ArecentsurveybyO’Reillyhighligh‐tedtheprivacyconcernsofcompaniesadoptingmachinelearningmodels,withmorethanhalfofcompaniesexperiencedwithAIMLcheckingforprivacyissues.

5

InthesameMITsurveymentionedpreviously,64%ofrespondentsnotethat“changesinregulationorgreaterregulatoryclarityondatasharing”isadevelopmentthatwouldbemostlikelytoleadtomoredatasharing.

Contemporaryprivacyregulations,suchastheUSHealthInsurancePortabilityandAccountabilityAct(HIPAA)andtheGeneralDataProtectionRegulation(GDPR)inEurope,imposeconstraintsorrequirementstousingpersonaldataforasecondarypurpose.Anexampleisarequirementtogetanadditionalconsentorauthoriza‐tionfromindividuals.Inmanycases,thisisnotpracticalandcanintroducebiasintothedatabecauseconsentersandnonconsentersdifferinimportantcharacteristics.

6

Datasynthesiscangivetheanalyst,ratherefficientlyandatscale,realisticdatatoworkwith.Giventhatsyntheticdatawouldnotbeconsideredidentifiablepersonaldata,privacyregulationswouldnotapply,andobligationsofadditionalconsenttousethedataforsec‐ondarypurposeswouldnotberequired.

7

ImprovingDataQuality

Giventhedifficultyingettingaccesstodata,manyanalyststrytojustuseopensourceorpublicdatasets.Thesecanbeagoodstartingpoint,buttheylackdiversityandareoftennotwellmatchedtotheproblemsthatthemodelsareintendedtosolve.Furthermore,open

4MITTechnologyReviewInsights,“TheGlobalAIAgenda:Promise,Reality,andaFutureofDataSharing”(March2020).

https://oreil.ly/FHg87

5BenLoricaandPacoNathan,TheStateofMachineLearningAdoptionintheEnterprise(O’Reilly).

6KhaledElEmam,etal.,“AReviewofEvidenceonConsentBiasinResearch,”AmericanJournalofBioethics13,no.4(2013):42–44.

https://oreil.ly/SiG2N.

7However,oneshouldfollowgoodpractices,suchasprovidingnoticetoindividualsabouthowthedataisusedanddisclosed,andhavingethicsoversightontheusesofdataandAIMLmodels.

TheBenefitsofSyntheticData|7

datamaylacksufficientheterogeneityforrobusttrainingofmodels.Forexample,theymaynotcapturerarecaseswellenough.

Sometimestherealdatathatexistsisnotlabeled.Labelingalargenumberofexamplesforsupervisedlearningtaskscanbetime-consuming,andmanuallabelingiserrorprone.Again,syntheticlabeleddatacanbegeneratedtoacceleratemodeldevelopment.Thesynthesisprocesscanensurehighaccuracyinthelabeling.

UsingSyntheticDataforExploratoryAnalysis

Analystscanusesyntheticdatamodelstovalidatetheirassumptionsanddemonstratethekindofresultsthatcanbeobtainedwiththeirmodels.Inthisway,thesyntheticdatacanbeusedinanexploratorymanner.Knowingthattheyhaveinterestingandusefulresults,theanalystscanthengothroughthemorecomplexprocessofgettingtherealdata(eitherraworde-identified)tobuildthefinalversionsoftheirmodels.

Forexample,ananalystwhoisaresearchercouldusetheirexplora-torymodelsonsyntheticdatatothenapplyforfundingtogetaccesstotherealdata,whichmayrequireafullprotocolandmultiplelev-elsofapprovals.Insuchaninstance,workwithsyntheticdatathatdoesnotproducegoodmodelsoractionableresultswouldstillbebeneficialbecauseanalystswouldhaveavoidedtheextraeffortrequiredtogetaccesstotherealdataforapotentiallyfutileanalysis.Anothervaluableuseofsyntheticdataisfortraininganinitialmodelbeforetherealdataisaccessible.Thenwhentheanalystgetstherealdata,theycanusethetrainedmodelasastartingpointfortrainingwiththerealdata.Thiscansignificantlyexpeditethecon-vergenceoftherealdatamodel(hencereducingcomputetime),andcanpotentiallyresultinamoreaccuratemodel.Thisisanexampleofusingsyntheticdatafortransferlearning.

UsingSyntheticDataforFullAnalysis

Avalidationservercanbedeployedtoruntheanalysiscodethatworkedonthesyntheticdataontherealdata.Ananalystwouldper-formalloftheiranalysisonthesyntheticdata,andthensubmitthecodethatworkedonthesyntheticdatatoasecurevalidationserverthathasaccesstotherealdata,asillustratedin

Figure1-1

.Becausethesyntheticdatawouldbestructuredinthesamewayastheorigi-naldata,thecodethatworkedonthesyntheticdatashouldwork

8|Chapter1:DefiningSyntheticData

directlyontherealdata.Theresultsarethensentbacktotheanalysttoconfirmtheirmodels.

Thisisnotintendedtobeaninteractivesystem.Theoutputfromthevalidationserverneedstobecheckedtoensurethatnorevealinginformationisbeingsentoutbythecodeoutput.Therefore,itisintendedtobeusedonceortwicebytheanalystattheveryendoftheiranalysis.Itdoesprovideawaytoprovideassurancetotheana-lyststhatthesynthesisresultsarereplicableontherealdata.

Figure1-1.Thesetupforavalidationserverusedtoexecutefinalcodethatproducedresultsonthesyntheticdata(adaptedfromReplica

AnalyticsLtd.,withpermission)

Whentheutilityofthesyntheticdataishighenough,theanalystscangetsimilarresultswiththesyntheticdataastheywouldhavewiththerealdata,andnovalidationserverisrequired.Insuchacase,thesyntheticdataplaystheroleofaproxyfortherealdata.Thisscenarioisplayingoutinmoreandmoreusecases:assynthesismethodsimproveovertime,thisproxyoutcomeisgoingtobecomemorecommon.

ReplacingRealDataThatDoesNotExist

Insomesituations,realdatamaynotexist.Theanalystmaybetry-ingtomodelsomethingcompletelynew,orthecreationorcollec-tionofarealdatasetfromscratchmaybecostprohibitiveorimpractical.Synthesizeddatacancoveredgeorrarecasesthataredifficult,impractical,orunethicaltocollectintherealworld.

Syntheticdatacanalsobeusedtoincreasetheheterogeneityofatrainingdataset,whichcanresultinamorerobustAIMLmodel.Forexample,unusualcasesinwhichdatadoesnotexistorisdifficulttocollectcanbesynthesizedandincludedinthetrainingdataset.In

LearningtoTrustSyntheticData|9

thatcase,theutilityofthesyntheticdataismeasuredintherobust‐nessincrementitgivestotheAIMLmodels.

Wehaveseenthatsyntheticdatacanplayakeyroleinsolvingaser‐iesofpracticalproblems.Onecriticalfactorfortheadoptionofdatasynthesis,however,istrustinthegenerateddata.Ithaslongbeenrecognizedthathighdatautilitywillbeneededforthebroadadop‐tionofdatasynthesismethods.

8

Thisisthetopicweturntonext.

LearningtoTrustSyntheticData

Initialinterestinsyntheticdatastartedintheearly’90swithpropos‐alstousemultipleimputationmethodstogeneratesyntheticdata.Imputationingeneralistheprocessofreplacingmissingdatavalueswithestimates.Missingdatacanoccur,forexample,inasurveyifsomerespondentsdonotcompleteaquestionnaire.

Accurateimputeddatarequirestheanalysttobuildamodelofthephenomenonofinterestbyusingtheavailabledataandthenusethatmodeltoestimatewhattheimputedvalueshouldbe.Tobuildavalidimputationmodel,theanalystneedstoknowhowthedatawillbeeventuallyused.Withmultipleimputation,youcreatemultipleimputedvaluestocapturetheuncertaintyintheseestimatedvalues.Thisprocesscanworkreasonablywellifyouknowhowthedatawillbeused.

Inthecontextofusingimputationfordatasynthesis,therealdataisaugmentedwithsyntheticdatabyusingthesametypeofimputationtechniques.Insuchacase,therealdataisusedtobuildanimputa‐tionmodelthatisthenusedtosynthesizenewdata.

Thechallengeisthatifyourimputationmodelsaredifferentfromtheeventualusesofthedata,theimputedvaluesmaynotbeveryreflectiveoftherealvalues,andthiswillintroduceerrorsinthedata.Thisriskofbuildingthewrongsynthesismodelhasledtohistoriccautionintheapplicationofsyntheticdata.

Morerecently,statisticalmachinelearningmodelshavebeenusedfordatasynthesis.Theadvantageofthesemodelsisthattheycancapturethedistributionsandcomplexrelationshipsamongthe

8JeromeP.Reiter,“NewApproachestoDataDissemination:AGlimpseintotheFuture(?),”CHANCE17,no.3(June2004):11–15.

https://oreil.ly/x89Vd

.

10|Chapter1:DefiningSyntheticData

variablesquitewell.Ineffect,theydiscovertheunderlyingmodelinthedataratherthanhavingthatmodelprespecifiedbytheanalyst.Andnowwithdeeplearningdatasynthesis,thesemodelscanbequiteaccurateinthattheycancapturemuchofthesignalinthedata—evensubtlesignals.

Therefore,wearegettingclosertothepointwherethegenerativemodelsavailabletodayareproducingdatasetsthatarebecomingquitegoodproxiesforrealdata.Therearealsowaystoassesstheutilityofsyntheticdatamoreobjectively.

Forexample,wecancomparetheanalysisresultsfromsyntheticdatawiththeanalysisresultsfromtherealdata.Ifwedonotknowwhatanalysiswillbeperformedonthesyntheticdata,arangeofpossibleanalysiscanbetriedbasedonknownexamplesofusesofthatdata.Oran“allmodels”evaluationcanbeperformedinwhichallpossiblemodelsarebuiltfromtherealandsyntheticdatasetsandcompared.

9

TheUSCensusBureauhas,atthetimeofwriting,decidedtolever‐agesyntheticdataforsomeofitsmostheavilyusedpublicdatasets,the2020decennialcensusdata.Foritstabulardatadisseminations,theagencywillcreateasyntheticdatasetfromthecollectedindividual-levelcensusdataandthenproducethepublictabulationsfromthatsyntheticdataset.Amixtureofformalandnonformalmethodswillbeusedinthesynthesisprocess.

10

Weprovideanover‐viewofthesynthesisprocessin

Chapter2

.This,arguably,demon‐stratesthelarge-scaleadoptionofdatasynthesisforoneofthemostcriticalandheavilyuseddatasetsavailabletoday.

Asorganizationsbuildtrustinsyntheticdata,theywillmovefromexploratoryanalysisusecases,totheuseofavalidationserver,andthentousingsyntheticdataasaproxyforrealdata.

Alegitimatequestioniswhataretheotherapproachesthatareavail‐abletodaytoaccessdataforAIMLpurposes,inadditiontodata

9AreviewofutilityassessmentapproachescanbefoundinKhaledElEmam,“Seven

WaystoEvaluat

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论