文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > IT/计算机 > 互联网 > ~Ad Click Prediction a View from the Trenches_2013
侵权投诉

~Ad Click Prediction a View from the Trenches_2013

online learning 的工程实现的论文

AdClickPrediction:aViewfromtheTrenches

H.BrendanMcMahan,GaryHolt,D.Sculley,MichaelYoung,

DietmarEbner,JulianGrady,LanNie,ToddPhillips,EugeneDavydov,

DanielGolovin,SharatChikkerur,DanLiu,MartinWattenberg,

ArnarMarHrafnkelsson,TomBoulos,JeremyKubica

Google,Inc.

mcmahan@google.com,gholt@google.com,dsculley@google.com

ABSTRACT

Predictingadclick–throughrates(CTR)isamassive-scalelearningproblemthatiscentraltothemulti-billiondollaronlineadvertisingindustry.WepresentaselectionofcasestudiesandtopicsdrawnfromrecentexperimentsinthesettingofadeployedCTRpredictionsystem.Theseincludeimprovementsinthecontextoftraditionalsupervisedlearn-ingbasedonanFTRL-Proximalonlinelearningalgorithm(whichhasexcellentsparsityandconvergenceproperties)andtheuseofper-coordinatelearningrates.

Wealsoexploresomeofthechallengesthatariseinareal-worldsystemthatmayappearat rsttobeoutsidethedomainoftraditionalmachinelearningresearch.Theseincludeusefultricksformemorysavings,methodsforas-sessingandvisualizingperformance,practicalmethodsforprovidingcon denceestimatesforpredictedprobabilities,calibrationmethods,andmethodsforautomatedmanage-mentoffeatures.Finally,wealsodetailseveraldirectionsthatdidnotturnouttobebene cialforus,despitepromis-ingresultselsewhereintheliterature.Thegoalofthispaperistohighlightthecloserelationshipbetweentheoreticalad-vancesandpracticalengineeringinthisindustrialsetting,andtoshowthedepthofchallengesthatappearwhenap-plyingtraditionalmachinelearningmethodsinacomplexdynamicsystem.

CategoriesandSubjectDescriptors

I.5.4[ComputingMethodologies]:PatternRecognition—Applications

Keywords

onlineadvertising,datamining,large-scalelearning

1.INTRODUCTION

Onlineadvertisingisamulti-billiondollarindustrythathasservedasoneofthegreatsuccessstoriesformachine

Permissiontomakedigitalorhardcopiesofpartorallofthisworkforpersonalorclassroomuseisgrantedwithoutfeeprovidedthatcopiesarenotmadeordistributedforpro torcommercialadvantageandthatcopiesbearthisnoticeandthefullcitationonthe rstpage.Copyrightsforthird-partycomponentsofthisworkmustbehonored.Forallotheruses,contacttheowner/author(s).

KDD’13,August11–14,2013,Chicago,Illinois,USA.Copyright2013ACM978-1-4503-2174-7/13/08...$15.00.

learning.Sponsoredsearchadvertising,contextualadvertis-ing,displayadvertising,andreal-timebiddingauctionshaveallreliedheavilyontheabilityoflearnedmodelstopredictadclick–throughratesaccurately,quickly,andreliably[28,15,33,1,16].Thisproblemsettinghasalsopushedthe eldtoaddressissuesofscalethatevenadecadeagowouldhavebeenalmostinconceivable.Atypicalindustrialmodelmayprovidepredictionsonbillionsofeventsperday,usingacorrespondinglylargefeaturespace,andthenlearnfromtheresultingmassofdata.

Inthispaper,wepresentaseriesofcasestudiesdrawnfromrecentexperimentsinthesettingofthedeployedsys-temusedatGoogletopredictadclick–throughratesforsponsoredsearchadvertising.Becausethisproblemsettingisnowwellstudied,wechoosetofocusonaseriesoftopicsthathavereceivedlessattentionbutareequallyimportantinaworkingsystem.Thus,weexploreissuesofmemorysavings,performanceanalysis,con denceinpredictions,cal-ibration,andfeaturemanagementwiththesamerigorthatistraditionallygiventotheproblemofdesigningane ec-tivelearningalgorithm.Thegoalofthispaperistogivethereaderasenseofthedepthofchallengesthatariseinrealindustrialsettings,aswellastosharetricksandinsightsthatmaybeappliedtootherlarge-scaleproblemareas.

2.BRIEFSYSTEMOVERVIEW

Whenauserdoesasearchq,aninitialsetofcandidateadsismatchedtothequeryqbasedonadvertiser-chosenkeywords.Anauctionmechanismthendetermineswhethertheseadsareshowntotheuser,whatordertheyareshownin,andwhatpricestheadvertiserspayiftheiradisclicked.Inadditiontotheadvertiserbids,animportantinputtotheauctionis,foreachada,anestimateofP(click|q,a),theprobabilitythattheadwillbeclickedifitisshown.

Thefeaturesusedinoursystemaredrawnfromavari-etyofsources,includingthequery,thetextoftheadcre-ative,andvariousad-relatedmetadata.Datatendstobeextremelysparse,withtypicallyonlyatinyfractionofnon-zerofeaturevaluesperexample.

Methodssuchasregularizedlogisticregressionareanat-ural tforthisproblemsetting.Itisnecessarytomakepredictionsmanybillionsoftimesperdayandtoquicklyupdatethemodelasnewclicksandnon-clicksareobserved.Ofcourse,thisdataratemeansthattrainingdatasetsareenormous.DataisprovidedbyastreamingservicebasedonthePhotonsystem–see[2]forafulldiscussion.

Becauselarge-scalelearninghasbeensowellstudiedinrecentyears(see[3],forexample)wedonotdevotesignif-

第1页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email