2024年4月15日发(作者：24小时精准天气预报)

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING

1

ASurveyonTransferLearning

SinnoJialinPanandQiangYangFellow,IEEE

Abstract—Amajorassumptioninmanymachinelearninganddataminingalgorithmsisthatthetrainingandfuturedatamustbe

r,inmanyreal-worldapplications,thisassumptionmaynothold.

Forexample,wesometimeshaveaclassificationtaskinonedomainofinterest,butweonlyhavesufficienttrainingdatainanother

domainofinterest,wherethelatterdatacases,

knowledgetransfer,ifdonesuccessfully,wouldgreatlyimprovetheperformanceoflearningbyavoidingmuchexpensivedatalabeling

ntyears,transfrveyfocuseson

categorizingandreviewingthecurrentprogressontransferlearningforclassification,survey,

wediscusstherelationshipbetweentransferlearningandotherrelatedmachinelearningtechniquessuchasdomainadaptation,multi-

tasklearningandsampleselectionbias,exploresomepotentialfutureissuesintransferlearning

research.

IndexTerms—TransferLearning,Survey,MachineLearning,DataMining.

✦

1I

NTRODUCTION

Dataminingandmachinelearningtechnologieshavealready

achievedsignificantsuccessinmanyknowledgeengineering

areasincludingclassification,,

[1],[2]).However,manymachinelearningmethodsworkwell

onlyunderacommonassumption:thetrainingandtestdataare

drawnfromthesamefeaturespaceandthesamedistribution.

Whenthedistributionchanges,moststatisticalmodelsneedto

manyrealworldapplications,itisexpensiveorimpossibleto

wouldbenicetoreducetheneedandefforttore-collectthe

cases,knowledgetransferortransfer

learningbetweentaskdomainswouldbedesirable.

Manyexamplesinknowledgeengineeringcanbefound

wheretransferlearningcantrulybebenefimple

isWebdocumentclassification[3],[4],[5],whereourgoal

istoclassifyagivenWebdocumentintoseveralpredefined

ampleintheareaofWeb-document

classification(,[6]),thelabeledexamplesmaybe

theuniversityWebpagesthatareassociatedwithcategory

informationobtainedthroughpreviousmanual-labelingefforts.

ForaclassificationtaskonanewlycreatedWebsitewherethe

datafeaturesordatadistributionsmaybedifferent,theremay

ult,wemaynotbe

abletodirectlyapplytheWeb-pageclassifierslearnedonthe

cases,itwould

behelpfulifwecouldtransfertheclassificationknowledge

intothenewdomain.

Theneedfortransferlearningmayarisewhenthedatacan

case,thelabeleddataobtainedin

onetimeperiodmaynotfollowthesamedistributionina

mple,inindoorWiFilocalization

DepartmentofComputerScienceandEngineering,HongKongUniversityof

ScienceandTechnology,ClearwaterBay,Kowloon,HongKong

Emails:{sinnopan,qyang}@

problems,whichaimstodetectauser’scurrentlocationbased

onpreviouslycollectedWiFidata,itisveryexpensiveto

calibrateWiFidataforbuildinglocalizationmodelsinalarge-

scaleenvironment,becauseauserneedstolabelalarge

r,the

WiFisignal-strengthvaluesmaybeafunctionoftime,device

trainedinonetimeperiod

orononedevicemaycausetheperformanceforlocation

estimationinanothertimeperiodoronanotherdevicetobe

cethere-calibrationeffort,wemightwishto

adaptthelocalizationmodeltrainedinonetimeperiod(the

sourcedomain)foranewtimeperiod(thetargetdomain),or

toadaptthelocalizationmodeltrainedonamobiledevice(the

sourcedomain)foranewmobiledevice(thetargetdomain),

asdonein[7].

Asathirdexample,considertheproblemofsentiment

classification,whereourtaskistoautomaticallyclassifythe

reviewsonaproduct,suchasabrandofcamera,intopositive

sclassificationtask,weneedto

firstcollectmanyreviewsoftheproductandannotatethem.

Wewouldthentrainaclassifieronthereviewswiththeir

hedistributionofreviewdata

amongdifferenttypesofproductscanbeverydifferent,to

maintaingoodclassificationperformance,weneedtocollect

alargeamountoflabeleddatainordertotrainthereview-

classifir,thisdata-

cethe

effortforannotatingreviewsforvariousproducts,wemay

wanttoadaptaclassificationmodelthatistrainedonsome

productstohelplearnclassificationmodelsforsomeother

cases,transferlearningcansaveasignificant

amountoflabelingeffort[8].

Inthissurveyarticle,wegiveacomprehensiveoverviewof

transferlearningforclassification,regressionandclustering

hasbeenalargeamountofworkontransferlearningfor

,

Digital Object Indentifier 10.1109/TKDE.2009.1911041-4347/$25.00 © 2009 IEEE

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING

[9],[10]).However,inthispaper,weonlyfocusontransfer

learningforclassification,regressionandclusteringproblems

g

thesurvey,wehopetoprovideausefulresourceforthedata

miningandmachinelearningcommunity.

ext

foursections,wefirstgiveageneraloverviewanddefine

brieflysurveythe

historyoftransferlearning,giveaunifieddefinitionoftransfer

learningandcategorizetransferlearningintothreedifferent

settings(giveninTable2andFigure2).Foreachsetting,we

reviewdifferentapproaches,

that,inSection6,wereviewsomecurrentresearchonthe

topicof“negativetransfer”,whichhappenswhenknowledge

ion7,

weintroducesomesuccessfulapplicationsoftransferlearning

andlistsomepublisheddatasetsandsoftwaretoolkitsfor

y,weconcludethearticlewith

adiscussionoffutureworksinSection8.

2O

VERVIEW

2.1ABriefHistoryofTransferLearning

Traditionaldataminingandmachinelearningalgorithmsmake

predictionsonthefuturedatausingstatisticalmodelsthatare

trainedonpreviouslycollectedlabeledorunlabeledtraining

data[11],[12],[13].Semi-supervisedclassification[14],[15],

[16],[17]addressestheproblemthatthelabeleddatamay

betoofewtobuildagoodclassifier,bymakinguseofa

largeamountofunlabeleddataandasmallamountoflabeled

ionsofsupervisedandsemi-supervisedlearning

forimperfectdatasetshavebeenstudied;forexample,Zhu

andWu[18]havestudiedhowtodealwiththenoisyclass-

eredcost-sensitivelearning

[19]whenadditionaltestscanbemadetofuturesamples.

Nevertheless,mostofthemassumethatthedistributionsof

erlearning,

incontrast,allowsthedomains,tasks,anddistributionsused

ealworld,we

mple,

wemayfindthatlearningtorecognizeapplesmighthelpto

rly,learningtoplaytheelectronicorgan

dyofTransfer

learningismotivatedbythefactthatpeoplecanintelligently

applyknowledgelearnedpreviouslytosolvenewproblems

damentalmotivation

forTransferlearninginthefieldofmachinelearningwas

discussedinaNIPS-95workshopon“LearningtoLearn”

1

,whichfocusedontheneedforlifelongmachine-learning

methodsthatretainandreusepreviouslylearnedknowledge.

Researchontransferlearninghasattractedmoreand

moreattentionsince1995indifferentnames:learningto

learn,life-longlearning,knowledgetransfer,inductivetrans-

fer,multi-tasklearning,knowledgeconsolidation,context-

sensitivelearning,knowledge-basedinductivebias,metalearn-

ing,andincremental/cumulativelearning[20].Amongthese,

:///courses/comp/dsilver/NIPS95LTL/

2

acloselyrelatedlearningtechniquetotransferlearningis

themulti-tasklearningframework[21],whichtriestolearn

multipletaskssimultaneouslyevenwhentheyaredifferent.

Atypicalapproachformulti-tasklearningistouncoverthe

common(latent)featuresthatcanbenefiteachindividualtask.

In2005,theBroadAgencyAnnouncement(BAA)05-29

ofDefenseAdvancedResearchProjectsAgency(DARPA)’s

InformationProcessingTechnologyOffice(IPTO)

2

gavea

newmissionoftransferlearning:theabilityofasystemto

recognizeandapplyknowledgeandskillslearnedinprevious

definition,transferlearningaims

toextracttheknowledgefromoneormoresourcetasksand

rasttomulti-task

learning,ratherthanlearningallofthesourceandtargettasks

simultaneously,transferlearningcaresmostaboutthetarget

esofthesourceandtargettasksarenolonger

symmetricintransferlearning.

Figure1showsthedifferencebetweenthelearning

wecansee,traditionalmachinelearningtechniquestrytolearn

eachtaskfromscratch,whiletransferlearningtechniquestry

totransfertheknowledgefromsomeprevioustaskstoatarget

taskwhenthelatterhasfewerhigh-qualitytrainingdata.

(a)TraditionalMachineLearning(b)TransferLearning

entLearningProcessesbetweenTraditional

MachineLearningandTransferLearning

Today,transferlearningmethodsappearinseveraltop

venues,mostnotablyindatamining(ACMKDD,IEEEICDM

andPKDD,forexample),machinelearning(ICML,NIPS,

ECML,AAAIandIJCAI,forexample)andapplicationsof

machinelearninganddatamining(ACMSIGIR,WWWand

ACLforexample)

3

.Beforewegivedifferentcategorizations

oftransferlearning,wefirstdescribethenotationsusedinthis

article.

2.2NotationsandDefinitions

Inthissection,weintroducesomenotationsanddefinitions

fall,wegivethedefinitions

ofa“domain”anda“task”,respectively.

Inthissurvey,adomainDconsistsoftwocomponents:a

featurespaceXandamarginalprobabilitydistributionP(X),

whereX={x

1

,...,x

n

}∈mple,ifourlearning

:///ipto/programs/tl/

arizealistofconferencesandworkshopswheretransfer

learningpapersappearinthesefewyearsinthefollowingwebpagefor

reference,/∼sinnopan/