講義課件說明random thoughts on kdd

Monday,September07,2015

ThoughtsonKDD2015

LastmonthIattendedKDD2015inbeautifulSydney,Australia.Forthosewhodon'tknow,KDDisthe internationalconferenceforappliedmachinelearning&datamining,andisoftenthevenueforsomeofthemostinterestingdata ysisresearchprojects.

DespiteconcernsthatKDD2015wouldbealetdownafterKDD2014wassuchagreatsuccessinNewYorkCity,overallKDD2015wasafantasticconference,withanexcellentlineupofinvitedspeakersandplentyofinterestingpapers.CongratulationsalsotomyPhDadvisorThorstenJoachims,whonotonlydidagreatjobasPCCo-Chair,butalsowastherecipientofaTestofTimeAwardforhisworkonOptimizingSearchEnginesusingClickthroughData.

DataScienceforScience

OneofthebiggestthemesatKDD2015wasapplyingdatasciencetosupportthesciences,whichissomethingthat'sbeenonmymindalotrecently.HughDurrant-Whitegaveagreatkeynoteonapplyingmachinelearningtodiscoveryprocessesingeologyandecology.Onethingthatjumpedoutofhistalkwashowchallengingitistodevelopmodelsthatareinterpretableto experts.Thisissueisamelioratedinhissettingsbecausehelargelyfocusedonspatialmodelswhichareeasiertovisualizeandinterpret.

SusanAtheygaveanotherkeynoteontheinterybetweenmachinelearningandcausalinferenceinevaluation,whichisanimportantissueforthesciencesaswell.Imustadmit,mostofthetalkwentovermyhead,buttherewassomeinterestingdebateafterthetalkaboutwhethercausalityshouldbethegoalorratherjustmore"robust"correlations(whateverthatmightmean).

IalsoreallyenjoyedtheData-DrivenSciencePanel,wherethedebategotquiteheatedattimes.Twoissuesinparticularstoodout.First,whatshouldbetheroleofmachinelearninganddataminingexpertsintheecosystemofdata-drivenscience?Onetheonehand,computerscientistshavehistoricallyhadalargeimpactbydevelosystemsandtformsthat awaylow-levelcomplexityandempowerusertobemoreproductive.However,howtoachievesuchasolutioninadata-richworldisamuessier(oratleastdifferent)typeofendeavor.Thereare,ofcourse,plentyofstartupsthataddressaspectsofthisproblem,butagenuinescalablesolutionforscienceremainselusive.

Asecondissuethatwasraisedwaswhethercomputationalresearchershavemademuchofadirectimpactonthesciences.Theparticulararea,raisedbyTinaEliassi-Rad,isthesocialsciences.Machinelearninganddatamininghavetakengreatinterestincomputationalsocialscienceviastudyinglargesocialnetworks.However,itisnotcleartowhatextentcomputationalresearchershavedirectlymadeanimpacttotraditionalsocialsciencefields.Ofcourse,thisissueistiedbacktowhattheroleofcomputationalresearchersshouldbe.Ontheonehand,manysocialscientistsdousetoolsmadebycomputationalpeople,sotheindirectimpactisquiteclear.Doesitreallymatterthattherehasn'tbeenmuchdirectimpact?

UpdateonMOOCs

DaphneKollergaveagreatkeynoteonthestateofMOOCsandCourserainparticular.ItseemsthatMOOCsnowadaysaremuchsmarterabouttheirconsumerbase,andhavediversifiedthewaytheydelivercontentandmeasuresuccessforawiderangeofstudents.Forexample,peoplenowunderstandmuchbetterthedifferentneedsofcollegeaspirants(whouseMOOCstosupplicanthighschool&collegeeducation)versusyoungprofessionals(whouseMOOCstogetaheadintheircareers)versusthoseseekingvocationalskills(whichisverypopularinlessdevelopedcountries).

OnestrikingomissionthatwaspointedoutduringtheQ&AwasthatMOOCshavemostlyabandonedthepre-collegedemographic,especiallybeforehighschool.Inretrospect,thisisnottoosurprising,inlargepartduetotheverydifferentrequirementsforprimaryandsecondaryeducationacrossdifferentstatesandschooldistricts.ButitdoesputadamperonthecurrentMOOCenthusiasm,sincemanyproblemswitheducationstartmuchearlierthancollege.

LessonsLearnedfromLarge-ScaleA/BTesting

RonKohavigaveakeynoteonlessonslearnedfromonlineA/Btesting.Themostinterestingaspectofhistalkwasjusthowwell-tunedtheexistingsystemsare.Onesymptomofahighlytunedsystemisthatit esverydifficulttointuitaboutwhethercertainmodificationswillincreaseordecreasetheperformanceofthesystem(orhavenoeffect).Forexample,hegavetheaudienceanumberofquestionstotheaudience,suchas:"Doesincreasingthedescriptionofthesponsoredadvertisementsleadtoincreasedoverallclicksonads?"Basically,theaudiencecouldnotguessbetterthanrandom.Sothemainlessonistobasicallytofollowthedataanddon'tbeto(emotionally)tiedtoyourownintuitionswhenitcomestooptimizinglargecomplexindustrialsystems.

Sports yticsWorkshop

Ico-organizedthe2ndworkshoponLarge-ScaleSportsytics.ItriedtogetmoreeSportsintotheworkshopthisyear,butalasfellabitshort.ThorstendidgiveaninterestingtalkthatusedeSportsdata,althoughthephenomenonhewasstudyingwasnotspecifictoeSports.Inmanyways,eSportsisanevenbettertestbedforsportsyticsthantraditionalsportsbecausegamereystrack

li llyeverything.

Withinthemoretraditionalsportsregimes,it'sclearth cesstodataremainsalargebottleneck.Manyprofessionalleaguesarehoardingtheirdatalikegold,butsadlydonothavetheexpertiseleveragethedataeffectively.ThesituationactuallyseemsbetterinEurope,whereaccesstotrackedsoccer(sorry,futbol)gamesarerelativelycommon.IntheUS,itseemslikethedataisonlyavailabletoaselectfewsportsyticscompaniessuchasSecondSpectrum.I'mhopefulthatthissituationwillchangeinthenearfutureasthevariousstakeholders emorecomfortablewiththeideathatit'snottherawdatathathasvalue,buttheprocessedartifactsbuiltontopofthatdata.

InterestingPapers

TherewereplentyofinterestingresearchpapersatKDD,ofwhichI'lljustlistafewthatIparticularlyliked.

ADecisionTreeFrameworkforSpatiotemporalSequencePrediction

byTaehwanKim,YisongYue,SarahTaylor,andIainMatthews

I'llstartwithashamelesspieceofself-advertising.IncollaborationwithDisneyResearch,wetrainedamodeltogeneratevisualspeech,i.e.,animatethelowerfaceinresponsetoaudioorphoneticinputs.Seethedemobelow:

Moredetailshere.

InsideJokes:IdentifyingHumorousCartoonCaptions

byDafnaShahaf,EricHorvitz,andRobertMankoff

ProbablythemostinterestingapplicationatKDDwasonstudyingtheanatomyofajoke.Whiletheresultsmaynotseemtoosurprisinginretrospect(e.g.,thepunchlineshouldbeatofthejoke),whatwasreallycoolwasthatthemodelcouldfyifonejokewasfunnierthananotherjoke(i.e.,rankjokes).

CinemaDataMining:TheSmellofFear

byJ?rgWicker,NicolasKrauter,BettinaDerstorff,ChristofSt?nner,EfstratiosBourtsoukidis,ThomasKlüpfel,JonathanWilliams,andStefanKramer

Thiswasacoolpaperthatstudiedhowtheexhaledorganicparticlesvaryinresponsetodifferentemotions.Theauthorsinstrumentedamovietheater'saircirculationsystemwithchemicalsensors,andfoundthatthechemicalsyouexhaleareindicativeofvariousemotionssuchasfearoramusement.Theauthorrepeatedlylamentedthefactthattheydidn'tdothisforanyeroticfi,andsotheydon'tknowwhatthecinematicchemicalsignatureofarousalwouldlooklike.

WhosupportedObamain2012?Ecologicalinferencethroughdistributionregression

bySethFlaxman,Yu-XiangWang,andAlexSmola

Thispaperpresentsanewsolutiontotheecologicalinferenceproblemofinferringindividuallevelpreferencesfromaggregatedata.Theprimarydatatestbedwerecounty-wiseelection esanddemographicdatathatreportedatadifferentgranularityoroverlay.Themainissueishowtoestimate,e.g.,femalepreferenceforoneialcandidate,usingjustthesekindsofaggregatedata.

Certifyingandremovingdisparateimpact

byMichaelFeldman,SorelleFriedler,JohnMoeller,CarlosScheidegger,andSureshVenkatasubramanian

Manypeopleassumethat,becausealgorithmsare"objective"thentheycan'tbebiasedordiscriminatory.Thisassumptionisinvalidbecausethedataorfeaturesthemselvescanbebiased(cf.thisinterviewwithCynthiaDwork).Theauthorsofthispaperproposeawaytodetect&removebiasinmachinelearningmodelsthatistailoredtotheUSlegaldefinitionofbias.Theworkis,ofcourse,preliminary,butthispaperwasarguablythemostthoughtprovokingoftheentireconference.

Edge-WeightedalizedPageRank:BreakingADecade-OldPerformanceBarrier

byWenleiXie,DavidBindel,AlanDemers,andJohannesGehrke

Thispaperproposesareductionapproachto alizedPageRankthatyieldsacomputationalboostbyseveralordersofmagnitude,thusallowing,forthefirsttime, alizePageRanktobecomputedatinctivespeeds.Thispaperwasalsotherecipientofthebestpaperaward.

PostedbyYisongYueat3:48PM

Labels:computerscience,machinelearning,science/technology

2comments:

BrendanO'Connorsaid...

whethercomputationalresearchershavemademuchofadirectimpactonthesciences--it'sagoodpointthatonlyasmallamountofcomputationalworkonostensiblesocialtopics