Gummadi Max Planck Institute for Software Systems MPISWS Germany Royal Veterinary College University of London United Kingdom CS Dept Federal University of Minas Gerais UFMG Brazil Abstract Directed links in social media could represent anything fro ID: 22754
Download Pdf The PPT/PDF document "Measuring User Inuence in Twitter The Mi..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
TheTwitterdatasetusedinthispaperconsistsof2billionfollowlinksamong54millionuserswhoproducedatotalof1.7billiontweets.Wereferreaderstoourprojectwebpagehttp://twitter.mpi-sws.org/foradetaileddescriptionofthedatasetandourdatasharingplan.Ourstudyprovidesseveralndingsthathavedirectimpli-cationsinthedesignofsocialmediaandviralmarketing:1)Analysisofthethreeinuencemeasuresprovidesabetterunderstandingofthedifferentrolesusersplayinsocialmedia.Indegreerepresentspopularityofauser;retweetsrepresentthecontentvalueofone'stweets;andmentionsrepresentthenamevalueofauser.Hence,thetopusersbasedonthethreemeasureshavelittleoverlap.2)Ourndingonhowinuencevariesacrosstopicscouldserveasausefultestforansweringhoweffectiveadver-tisementinTwitterwouldbeifoneistoemployinuentialusers.Ouranalysisshowsthatmostinuentialusersholdsignicantinuenceoveravarietyoftopics.3)Ordinaryuserscangaininuencebyfocusingonasin-gletopicandpostingcreativeandinsightfultweetsthatareperceivedasvaluablebyothers,asopposedtosimplyconversingwithothers.Thesendingsprovidenewinsightsforviralmarketing.Therstndinginparticularindicatesthatindegreealonerevealslittleabouttheinuenceofauser.ThishasbeencoinedthemillionfollowerfallacybyAvnit(Avnit2009),whopointedtoanecdotalevidencethatsomeusersfollowotherssimplyforetiquette—it'spolitetofollowsomeonewho'sfollowingyou—andmanydonotreadallthebroad-casttweets.WehaveempiricallydemonstratedthathavingamillionfollowersdoesnotalwaysmeanmuchintheTwitterworld.Instead,weclaimthatitismoreinuentialtohaveanactiveaudiencewhoretweetsormentionstheuser.DeninginuenceonTwitterWestartbyreviewingstudiesofdiffusionofinuenceandrelatedworkoninuencepropagationonTwitter.BackgroundThereareanumberofconictingideasandtheoriesabouthowtrendsandinnovationsgetadoptedandspread.Thetraditionalviewassumesthataminorityofmem-bersinasocietypossessqualitiesthatmakethemexcep-tionallypersuasiveinspreadingideastoothers.Theseex-ceptionalindividualsdrivetrendsonbehalfofthemajor-ityofordinarypeople.Theyarelooselydescribedasbe-inginformed,respected,andwell-connected;theyarecalledtheopinionleadersinthetwo-stepowtheory(KatzandLazarsfeld1955),innovatorsinthediffusionofinnovationstheory(Rogers1962),andhubs,connectors,ormavensinotherwork(Gladwell2002).Thetheoryofinuentialsisintuitiveandcompelling.Byidentifyingandconvincingasmallnumberofinuentialindividuals,aviralcampaigncanreachawideaudienceatasmallcost.Thetheoryspreadwellbeyondacademiaandhasbeenadoptedinmanymarketingbusinesses,e.g.,RoperASWandTremor(Gladwell2002;BerryandKeller2003).Incontrast,amoremodernviewofinformationowem-phasizestheimportanceofprevailingculturemorethantheroleofinuentials.Someresearchershavereasonedthatpeopleinthenewinformationagemakechoicesbasedontheopinionsoftheirpeersandfriends,ratherthanbyinu-entials(DomingosandRichardson2001).Theseresearchersarguedthatdirectmarketingthroughinuentialswouldnotbeasprotableasusing“network”-basedadvertisingsuchascollaborativeltering.Thetraditionalinuentialstheoryhasalsobeencriticizedbecauseitsinformationowprocessdoesnottakeintoac-counttheroleofordinaryusers.Inordertocomparetheroleofinuentialsandordinaryusers,researchershavede-velopedaseriesofsimulations,inwhichinformationowsfreelybetweenusers,andauseradoptsaninnovationwhenheisinuencedbymorethanathresholdofthesamplepop-ulation(WattsandDodds2007).Inuentialsweredenedasthoseinthetop10%ofinuencedistribution.Thesimu-lationshowedthatinuentialsinitiatedmorefrequentandlargercascadesthanaverageusers,buttheywereneithernecessarynorsufcientforalldiffusions,assuggestedinthetraditionaltheory.Moreover,inhomogeneousnetworks,inuentialswerenomoresuccessfulinrunninglongcas-cadesthanordinaryusers.Thismeansthatatrend'ssuccessdependsnotonthepersonwhostartsit,butonhowsuscep-tiblethesocietyisoveralltothetrend(Watts2007).Infact,atrendcanbeinitiatedbyanyone,andiftheenvironmentisright,itwillspread.Therefore,Wattsdubbedearlyadoptersoropinionleaders“accidental”inuentials.Theabovecompetingideashaveremainedashypothesesforseveralreasons.Firstisthelackofdatathatcouldbeusedtoempiricallytestthem.Althoughthereexistahandfulofempiricalstudiesonword-of-mouthinuence(Leskovec,Adamic,andHuberman2007;Cha,Mislove,andGummadi2009),noworkhasbeenconductedontherelativeorderofinuenceamongindividuals.Asecondissueistheva-rietyofwaysthatinuencehasbeendened(Watts2007;Goyal,Bonchi,andLakshmanan2010).Ithasbeenunclearwhatexactlyinuencemeans.Finally,decadeshavepassedsincetheinuentialstheoryappeared.Evenifthetheorywasreasonablyaccuratewhenitwasproposed,thingshavechangedandnowwehavemuchmorevariabilityintheowofinuence.Inparticular,onlinecommunitieshavebecomeasignicantwaywereceivenewinformationandinuenceinsuchcommunitiesneedstobeexplored.Inthispaper,weinvestigatethenotionofinuenceusingalargeamountofdatacollectedfromapopularsocialmedium,Twitter.MeasuringinuenceonTwitterTheMerriam-Websterdictionarydenesinuenceas“thepowerorcapacityofcausinganeffectinindirectorintan-gibleways.”Despitethelargenumberoftheoriesofinu-enceinsociology,thereisnotangiblewaytomeasuresuchaforcenoristhereaconcretedenitionofwhatinuencemeans,forinstance,inthespreadofnews.Inthispaper,weanalyzetheTwitternetworkasanewsspreadingmediumandstudythetypesanddegreesofinu-encewithinthenetwork.Focusingonanindividual'spo-tentialtoleadotherstoengageinacertainact,wehighlight trarymonotonicfunctioncoulddescribetherelationshipbe-tweentwovariables,withoutmakinganyotherassumptionsabouttheparticularnatureoftherelationshipbetweenthevariables.Ourinclusiveandcompletedatasetguaranteesre-liabilityofthecorrelationestimates.Thecloseristo+1or�1,thestrongerthelikelycorrelation.Aperfectpositivecorrelationis+1andaperfectnegativecorrelationis�1.ComparingthreemeasuresofuserinuenceToseewhatkindsofusersarethemostinuential,wevisitedtheTwitterpagesofthetop-20usersbasedoneachmeasure.ThetopinuentialsThemostfollowedusersspanawidevarietyofpublicguresandnewssources.Theywerenewssources(CNN,NewYorkTimes),politicians(BarackObama),athletes(ShaquilleO'Neal),aswellascelebritieslikeactors,writers,musicians,andmodels(AshtonKutcher,BritneySpears).Asthelistsuggests,indegreemeasureisusefulwhenwewanttoidentifyuserswhogetlotsofat-tentionfromtheiraudiencethroughone-on-oneinteractions,i.e.,theaudienceisdirectlyconnectedtoinuentials.Themostretweeteduserswerecontentaggregationser-vices(Mashable,TwitterTips,TweetMeme),businessmen(GuyKawasaki),andnewssites(TheNewYorkTimes,TheOnion).Theyaretrackersoftrendingtopicandknowledge-ablepeopleindifferentelds,whomotherusersdecidetoretweet.Unlikeindegree,retweetsrepresentinuenceofauserbeyondone'sone-to-oneinteractiondomain;populartweetscouldpropagatemultiplehopsawayfromthesourcebeforetheyareretweetedthroughoutthenetwork.Further-more,becauseofthetightconnectionbetweenusersassug-gestedinthetriadicclosure(Granovetter1973),retweetinginasocialnetworkcanserveasapowerfultooltoreinforceamessage—forinstance,theprobabilityofadoptinganin-novationincreaseswhennotonebutagroupofusersrepeatthesamemessage(WattsandDodds2007).Themostmentionedusersweremostlycelebrities.Ordi-naryusersshowedagreatpassionforcelebrities,regularlypostingmessagestothemormentioningthem,withoutnec-essarilyretweetingtheirposts.Thisindicatesthatcelebritiesareofteninthecenterofpublicattentionandcelebritygos-sipisapopularactivityamongTwitterusers.Ifretweetsrepresentacitationofanotheruser'scon-tent,mentionsrepresentapublicresponsetoanotheruser'stweet—thefocusofatweetisoncontentforretweets,whilethefocusisonthereplieduserformentions.Thiscanbeconrmedfromtheusageofconventionsintweets:92%oftweetsthathadaRTorviamarkercontainedaURLand97%ofthemalsocontainedthe@usernameeld.Thismeansthatretweetsareaboutthecontent(indicatedbytheembeddedURL)andthatpeopletypicallycitetheauthen-ticsourcewhentheyretweet.However,fewerthan30%oftweetsthatwereclassiedasmentionscontainedanyURL,indicatingthatamentionismoreidentity-driven.Acrossallthreemeasures,thetopinuentialsweregener-allyrecognizablepublicguresandwebsites.Interestingly,wesawmarginaloverlapinthesethreetoplists.Thesetop-20listsonlyhad2usersincommon:AshtonKutcherandPuffDaddy.Thetop-100listsalsoshowedmarginalover- Figure1:Venndiagramofthetop-100inuentialsacrossmeasures:Thechartisnormalizedsothatthetotalis100%.lap,asshowninFigure1,indicatingthatthethreemeasurescapturedifferenttypesofinuence.RelativeinuenceranksInordertoinvestigatehowthethreemeasurescorrelate,wecomparedtherelativeinuenceranksofall6millionusers(Table1).Weseeamoderatelyhighcorrelation(above0.5)acrossallpairs.However,thehighcorrelationappearstobeanartifactofthetiedranksamongtheleastinuentialusers,e.g.,manyoftheleastcon-nectedusersalsoreceivedzeroretweetandmention.Toavoidthisbias,wefocusedonthesetofrelativelypopu-larusers.Weconsideredusersinthetop10thand1stper-centilesbasedonindegree,inthehopethatuserswhogetretweetedormentionedmusthavesomefollowers.Table1:Spearman'srankcorrelationcoefcients CorrelationAllTop10%Top1% Indegreevsretweets0.5490.1220.109Indegreevsmentions0.6380.2860.309Retweetsvsmentions0.5800.6380.605 Afterthislteringstep,thetopusersshowedastrongcorrelationintheirretweetinuenceandmentioninuence.Samplingthetopusersbasedonretweetsormentionsleadstosimilarresults.Thismeansthat,ingeneral,userswhogetmentionedoftenalsogetretweetedoften,andviceversa.In-degree,however,wasnotrelatedtotheothermeasures.Weconcludethatthemostconnectedusersarenotnecessarilythemostinuentialwhenitcomestoengagingone'saudi-enceinconversationsandhavingone'smessagesspread.DiscussionofmethodologyNormalizingretweetsandmentionsbytotaltweetswouldyieldadifferentmeasureofinuence,whichmighthaveledtoverydifferentresults.Whenwetriednormalizingthedata,weidentiedlocalopinionleadersasthemostinuential.However,normal-izationfailedtorankuserswiththehighestsheernumberofretweetsasinuential.Therefore,inthispaper,weusethesheernumberofretweetsandmentionswithoutnormalizingthesevaluesbythetotaltweetsofauser.Othermeasuressuchasthenumberoftweetsandout-degree(i.e.,thenumberofpeopleauserfollows)werenotfoundtobeuseful,becausetheyidentiedrobotsandspam-mersasthemostinuential,respectively.Therefore,wedonotusethesemeasures. Table2:Summaryinformationofthethreemajortopicseventsstudied Topic PeriodKeywords UsersTweetsAudience Iran Jun11—Aug10#iranelection,namesofpoliticians 302,1301,482,03822,177,836Swine May3—July2Mexicou,H1N1,swine 239,329495,82520,977,793Jackson Jun25—Aug24MichaelJackson,#mj 610,2131,418,35623,550,211 Finally,wecalculatedindegreebasedonthesnapshotofthenetworkatthetimeofcrawlingin2009,becausewedonotknowthetimewheneachfollowlinkwasformed.Forthecalculationofretweetsandmentions,however,weusedlongitudinaldata(i.e.,sincethebeginningoftheTwit-terservicein2006).Thisdifferencecouldhaveresultedinaweakercorrelationbetweenindegreeandtheotherinu-encemeasures(Table1).Nevertheless,nearlythreequartersofTwitterusersjoinedin2009,suggestingthattheeffectonthecorrelationwouldhavebeenminimal.Doesinuenceholdacrossdifferenttopics?HavingdenedthreedifferentmeasuresofinuenceonTwitter,wenowinvestigatewhetherauser'sinuencevariesbytopicgenres.Inordertoinvestigatethis,werstneedtonduserswhotweetedaboutadiversesetoftopics.MethodologyforidentifyingtargettopicsTondasmanyuserstomonitoraspossible,wepickedthreeofthemostpopulartopicsin2009thatwereconsidereden-gagingandrevolutionaryinTwitter2:theIranianpresiden-tialelection,theoutbreakoftheH1N1inuenza,andthedeathofMichaelJackson.Whileallthreetopicswerepopu-lar,theyspanpolitical,health,andsocialgenres.Toextracttweetsrelevanttotheseevents,werstidentiedthesetofkeywordsdescribingthetopicsbyconsultingnewswebsitesandinformedindividuals.Usingtheselectedkeywords,wethenidentiedrelevanttweetsbysearchingforthekeywordsinthetweetdataset.Table2displaysthekeywordsandthetotalnumberofusersandtweetsforeachtopic.Wefocusedonaperiodof60daysstartingfromonedaypriortothestartofeachevent.Welimitedthestudydurationbecausepopularkeywordsweretypicallyhijackedbyspammersaftercertaintime.Thetablealsoshowsthetotalnumberofuserswhoreceivedanytweetonthetopic(termedaudience).Eachtopicreachedanaudienceofover20million,indicatingthatover40%ofusersinTwitterwereawareofatleastoneofthethreetopics.Amongtheuserswhotweetedaboutanyofthesetopics,fewerthan2%discussedallthreetopics.Although2%isasmallfraction,thissetcontains13,219users,whichisalargeenoughsamplesizeforstatisticalanalysis.Theseusersweregenerallywellconnected;theyhadonaverage2,037follow-ers,andtogetherreachedanaudienceof16million.Fur-thermore,noneoftheseuserswerededicatedtothesethreetopics;nouserhaddedicatedmorethan60%oftheirtweetstothesetopics.Thismeansthatthesetof13,219usersisenthusiasticaboutsharingthoughtsonpopularnewstopics 2ThetopTwittertrendsidentiedbytheTwitterResearchteamarelistedathttp://tinyurl.com/yb4965e.fromdiversegenres.Thesepropertiesmakethisgroupidealforstudyinghowauser'sinuencevariesacrossvastlydif-ferenttopicgenres.DistributionoftheinuenceranksTogetameasureofinuenceforagiventopic,wecountonlytheretweetsandmentionsauserspawnedonthegiventopic.Becauseindegreeisinvariantacrosstopics,wedonotusethemeasureinthisanalysis.Beforeinvestigatingthedynamicsofinuenceacrosstop-ics,werstneedtounderstandthehigh-levelcharacteris-ticsofuserinuence,suchasthedegreetowhichusers'inuencecandiffer.Figure2displaystheinuenceranksofusersbasedonretweetsandmentionsacrosstheIran,Swine,andJacksontopics.Theinuenceranksarecalcu-latedforeachtopicandauser'srankmaydifferfordifferenttopics.Theplotsshowastraightlineonalog-logplot,apropertythatisreferredtoasthepower-lawcharacteristic.Thepower-lawpatternisindicativeofthefactthatusers'degreeofinuencecandifferbyordersofmagnitude:thetopinuentialswereretweetedormentioneddisproportion-atelymoretimesthanthemajorityofusers.Thissuggeststhatutilizingtopinuentialshasagreatpotentialpayoffinmarketingstrategy. (a)Retweetinuenceranks (b)MentioninuenceranksFigure2:Distributionofuserranksforagiventopic Variationofauser'sinuenceacrosstopicsNext,inordertoexaminehowvolatileauser'srankisacrossdifferenttopics,wecomparetherelativeorderofinuenceranksacrosstopicsinTable3.Forthesamereasonweprevi-ouslydiscussedinTable1,weignoretheleastpopularuserswhohavetiedranksandfocusonthesetofrelativelypop-ularusers,asmeasuredbyindegree.Thetop10thand1stpercentilesincluded1,322and132users,respectively.Table3:Spearman'srankcorrelationcoefcientsovertopics RetweetMentionsTopicsTop10%1%Top10%1% Iranvs.Swine0.540.620.590.68Iranvs.Jackson0.480.540.590.63Swinevs.Jackson0.550.500.800.68Therankcorrelationisgenerallyhigh(above0.5)andgetsstrongerforthetop1%ofusers.Mentionsshowanevenstrongercorrelationacrosstopicsthanretweets.Thismeansthatapopularuserwhoisgoodatspawningmentionsfromotherscandosooverawiderangeoftopics,moreeasilythanwhensheisretweetedoverdiversetopics.Amongtopicpairs,SwineandJacksonshowedthehighestcorrelationforthetop10%ofwell-connectedusersforboththeretweetandmentioninuence.Thisisperhapsduetothemoresocialnatureofthesetwoevents,whichdiffersfromtheIrantopicwherespecialinterestgroupslikepoliticiansandbloggersplayedamajorrole.Giventhattherankcorrelationgetsstrongerforuserswithhighindegree,weinvestigatedhowwidelytheranksofthetoplisteduserschangebytopic.Figure3showstheretweetranksofthetopusersintheIrantopicagainsttherelativeretweetranksofthesameusersintheSwineandJacksontopics.Itisclearfromthegurethatthetop5usersintheIrantopicretainedtheirrelativeranksintheothertwotopics.TheseuserswereMashable,CnnBrk,TweetMeme,Time,andBreakingNews,whowereallinthecategoryofauthoritativenewssourcesandcontenttrackers.Themen-tioninuencealsoshowedasimilartrend(notshownhere);mostinuentialsrankedconsistentlyhighamongstdifferenttopics.Andwehaveobservedthistrendnotonlyforthehyper-inuentials:moderateinuentialslikeopinionlead-ersandevangelistsalsohadconsistentinuenceranksoverdiversetopics,asshowninTable3.Ourndingsaboutthehighlyskewedabilityofuserstoinuenceothers(Figure2)andthestrongcorrelationinauser'sinuencerankacrossdifferenttopics(Table3)to-getherleadtotwointerestingconclusions.First,mostinu-entialusersholdsignicantinuenceoveravarietyoftop-ics.Thismeansthatlocalopinionleadersandhighlypopulargurescouldindeedbeusedtospreadinformationoutsidetheirareaofexpertise.Infact,newadvertisementcampaignshaverecentlybeenlaunchedthatinsertadvertisementlinksintoapopularperson'stweet(Fiorillo2009).Second,thepower-lawtrendinthedifferenceamonginuenceofindi-vidualsindicatesthatitissubstantiallymoreeffectivetotar-getthetopinuentialsthantoemployamassivenumberofnon-popularusersinordertokickstartaviralcampaign. Figure3:Users'retweetinuenceranksovertopicsTheriseandfallofinuentialsovertimeManyfactors—social,political,andeconomical—affectpopularityandinuenceofindividualsandorganizations.Inonlinesocialmedia,suchdynamicsisfacilitatedbyeasyen-tryofcompetition.Itonlycosts140characterstogenerateatweetforanyuser.Likewise,itwillbechallengingforin-uentialstomaintaintheirstatuswhenmanyemerginglocalopinionleadersandevangelistsenterthearena.Hereweanalyzethedynamicsofindividual'sinuenceovertimeintwoways.First,wetrackthepopularityoftopinuentialsoveralongtermperiodandcheckhowwelltheymaintaintheirranks.Second,wefocusonuserswhoin-creasedtheirinuenceinaspecictopicoverashorttimeperiod,inordertounderstandwhatbehaviorsmakeordinaryindividualsinuential.MaintainingengagementofthetopinuentialsOutofall6millionactiveusersintheTwitternetwork,wepickedthetop100usersbasedoneachofthethreemea-sures:indegree,retweets,andmentions.WeusedallthetweetseverpostedonTwitterinidentifyingtheseinuen-tials.DuetotheoverlapwediscussedinFigure1,therewere233distinctusers,whomwecallall-timeinuentials.Inordertoseehowtheinuenceoftheseall-timeinuen-tialsvariedovertime,wetrackedtheirinuencescoresoveran8monthperiodfromJanuarytoAugust2009.Togetatime-varyingmeasureofinuence,wecountedthenumberofretweetsandmentionstheall-timeinuen-tialsspawnedinevery15daysoverthe8monthperiod.BecauseweonlyknowtheindegreeinformationbasedonthenalsnapshotoftheTwitternetwork,wedonotusethismeasure.Foreachuser,wecomputedasingleexplanatoryvariableP:theprobabilitythatarandomtweetpostedonTwitterduringa15dayperiodisaretweet(oramention)ofthatuser.NormalizingbythetotalnumberoftweetspostedonTwitterisessentialtocanceloutanyvariableeffectonthedataandallowstheunderlyingcharacteristicsofthedatasetstobecompared.Forinstance,becausetheTwitternetworkquadrupledovertimeintermsoftheregisteredusers,theto-talvolumeoftweetsmerelyincreasedovertime.Hence,ifwedidn'tnormalizetheresults,thetrendwouldn'tbeinter-esting.Googlesimilarlynormalizesthedatawhenanalyzingtheirsearchtrends(Ginsbergetal.2009).Figure4displaysthetimeevolutionofthenormalizedretweetsandmentionsofthe233all-timeinuentialusers. (a)Retweetinuence (b)MentioninuenceFigure4:Thetemporalevolutionofretweetsandmentionsfortheall-timeinuentialusers.Foreachdatapoint,theerrorbarsarecenteredontheaverageretweet(ormention)probability,andtheyextendupanddownbytwostandarddeviations.AlthoughthevaluesofPappearsmall,theyaccountforalargevolumeofretweetsandmentions:asinglemostpopularinuentialuserspawnedupto20,000retweetsand50,000mentionsoverarandom15dayperiod.Inordertocapturethetrendinmoredetail,weclassiedtheinuen-tialsintothreegroupsbasedonindegree:thetop10users,whoweremostlythemainstreamnewssources;thenext90users,whoweremainlycelebrities;andtherest,whowereamixedgroupofpublicguresandopinionleadersthatcom-petedwiththetraditionalmassmedia.TheevolutionoftheretweetprobabilityinFigure4(a)showsthatallthreegroupsmildlyincreasedtheirinuenceovertime.Largeerrorrangesforthetop10usersindicateamuchwidervariabilityinthepopularityofthisgroup.Thegrowth,however,ismarginalforallthreegroups.Wecon-jecturethatthemarginalincreaseisduetothelimitednum-beroftweetsuserspostaday.Broadcastingtoomanytweetsputsevenpopularusersatariskofbeingclassiedasspam-mers.Hence,Twitterusersshouldmoderatethenumberofbroadcasttweetsinordertoavoidcrossingtheirfollowers'informationprocessinglimit.TheevolutionofthementionprobabilityinFigure4(b)showsdistinctpatternsforthethreegroups.Thetop10usersfellinpopularityovertime;thenextgrouphadconsis-tentinuencescoresovertime;andthelastgroup(theleastconnectedamongthethree)increasedtheirpopularityovertime.Usersinthelastgroup,surprisingly,spawnedonaver-agemorementionsthanthetop100users.Whilethistrendiscounter-intuitiveatrst,thedifferencesbetweenmentionsandretweetscanexplainthetrend.Themainstreamnewsorganizationsintherstgroupareretweetedthemost,buttheyarenotmentionedthemost.Thisisbecausetheirnamescomeupmostlywhentheircon-tentgetretweeted;itishardformediasourcestoengageuserswiththeiridentitiesalone.Thesecondgroup,com-prisedofcelebrities,ismoreoftenmentionedthanretweetedbecauseoftheirnamevalue.Theirtweetsalsogetretweeted,wheninuenceistransferredacrosstopics(Table3).Evan-gelistsinthelastgroupsuccessfullyincreasedtheirinu-ence.Whilemanyfactorscouldexplainthisphenomenon,ourmanualinspectionrevealedthattheseusersputsigni-canteffortsinconversingwithothers(e.g.,replyingtotheiraudience).Inasense,theyneedself-advertisementthemost,becausemassmediaandcelebritieshavemanyotheron-andoff-linechannelsoftopromotethemselves.Whileourndingsprovideaninterestingviewofhowdif-ferentgroupsofpeoplemaintaintheirpopularity,weshouldalsoemphasizethatouranalysisisinretrospectbutnotcausal.Thesendingsarebasedonthesetofuserswhoultimatelybecamepopular.Wealsomentionthatallinu-entialsputeffortsinpostingcreativeandinterestingtweets,asshownfromthehighcorrelationintheirretweetranks.RisinginuenceoftheordinaryusersFinallyweexaminedtheuserswhoincreasedtheirinuenceoverashorttimeperiodtounderstandwhatbehaviorsmakeordinaryusersinuential.Wefocusedonthesetofuserswhotalkedaboutonlyonenewstopic,outofthethreenewstopicsinTable2,andpickedthetop20usersforeachnewstopicbasedonindegree.Wecalltheseuserstopicalinu-entials.Thetopicalinuentialsincludeddedicatedaccountslikeiranbaan,oxfordgirl,andTM Outbreakwhosuddenlybecamepopularoverthecourseoftheevent.Theseuserswereliterallyunheard-ofatthebeginningofthenewseventsanddidn'treceiveanyretweetsormentionspriortotherel-evantnewsevent.Thelistalsoincludeduserslikekevinrose(thefounderofdigg.com)and106andpark(BET.com'smu-sicvideosite)whowerealreadypopularonaspecictopic,andusedthenewseventstoextendtheirpopularity.Wetrackedtheinuencescoresofthe60topicalinu-entialsoverthe8monthin2009.Again,wecomputedthevariablePastheprobabilitythatarandomtweetpostedina15dayperiodisaretweet(oramention)tothatuser.Fig-ure5displaysthetemporalevolutionofinuenceforthetopicalinuentials.Overall,theinuencescoresaremuchlowerthanthatoftheglobalinuentialsinFigure4.Thisisexpectedsincetopicalinuentialshad3to180timesfewerfollowersthantheglobalinuentials.InuentialsonSwineuexperiencedrelativelystablein-uencescoresforbothretweetsandmentions.Thisisbe-causenosingledayinvolvedacatastrophiceventduetoSwineu.InuentialsontheIranelectionincreasedtheir (a)Retweetinuence (b)MentioninuenceFigure5:Thetemporalevolutionofretweetsandmentionsforthetopicalinuentialusersretweetinuencesubstantiallyduringthepeakoftheelec-tioninJuneandJuly,buttheydidnotspawnmanymen-tions.Ourdatashowsthattheseinuentialsactivelyspreadinformationaboutprotestsandcontroversialnews.Inuen-tialsonMichaelJacksonexperiencedamildincreaseintheirretweetinuenceinJune,whereastheirmentioninuencesurgedduringthesametimeperiod.Ourmanualinspectionrevealedthatuserswholimittheirtweetstoasingletopicshowedthelargestincreaseintheirinuencescores.ConclusionsThispaperanalyzedtheinuenceofTwitterusersbyem-ployingthreemeasuresthatcapturedifferentperspectives:indegree,retweets,andmentions.Wefoundthatindegreerepresentsauser'spopularity,butisnotrelatedtootherim-portantnotionsofinuencesuchasengagingaudience,i.e.,retweetsandmentions.Retweetsaredrivenbythecontentvalueofatweet,whilementionsaredrivenbythenamevalueoftheuser.SuchsubtledifferencesleadtodissimilargroupsofthetopTwitterusers;userswhohavehighinde-greedonotnecessarilyspawnmanyretweetsormentions.Thisndingsuggeststhatindegreealonerevealsverylittleabouttheinuenceofauser.Focusingonretweetsandmentions,westudiedthedy-namicsofinuenceacrosstopicsandtime.Ourspatialanal-ysisshowedthatmostinuentialuserscanholdsignicantinuenceoveravarietyoftopics.ThetopTwitterusershadadisproportionateamountofinuence,whichwasin-dicatedbyapower-lawdistribution.Ourtemporalanaly-sisidentiedhowdifferenttypesofinuentialsinteractwiththeiraudience.Mainstreamnewsorganizationsconsistentlyspawnedahighlevelofretweetsoverdiversetopics.Incon-trast,celebritieswerebetteratinducingmentionsfromtheiraudience.Thisisbecausethenamevalueofthementionin-uentialshelpedthemgetresponsesfromothers,ratherthananyinherentvalueinthecontenttheyposted.Finally,wefoundthatinuenceisnotgainedsponta-neouslyoraccidentally,butthroughconcertedeffort.Inor-dertogainandmaintaininuence,usersneedtokeepgreatpersonalinvolvement.Thiscouldmeanthatinuentialusersaremorepredictablethansuggestedbytheory(Watts2007),sheddinglightonhowtoidentifyemerginginuentialusers.ReferencesAvnit,A.2009.TheMillionFollowersFallacy,InternetDraft,PravdaMedia.http://tinyurl.com/nshcjg.Berry,J.,andKeller,E.2003.TheInuentials:OneAmericaninTenTellstheOtherNineHowtoVote,WheretoEat,andWhattoBuy.FreePress.Buck,W.1980.TestsofSignicanceforPoint-BiserialRankCor-relationCoefcientsinthePresenceofTies.BiometricalJournal22(2):153–158.Cha,M.;Mislove,A.;andGummadi,K.P.2009.AMasurement-DrivenAnalysisofInformationPropagationintheFlickrSocialNetwork.InWWW.Domingos,P.,andRichardson,M.2001.MiningtheNetworkValueofCustomers.InACMSIGKDD.Fiorillo,J.2009.TwitterAdvertising:Pay-Per-Tweet,Wiki-nomics.http://tinyurl.com/d89uu9.Ginsberg,J.;Mohebbi,M.H.;Patel,R.S.;Brammer,L.;Smolin-ski,M.S.;andBrilliant,L.2009.DetectingInuenzaEpidemicsUsingSearchEngineQueryData.Nature457.Gladwell,M.2002.TheTippingPoint:HowLittleThingsCanMakeaBigDifference.BackBayBooks.Goyal,A.;Bonchi,F.;andLakshmanan,L.V.S.2010.LearningInuenceProbabilitiesInSocialNetworks.InACMWSDM.Granovetter,M.1973.TheStrengthofWeakTies.AmericanJournalofSociology78(6).Katz,E.,andLazarsfeld,P.1955.PersonalInuence:ThePartPlayedbyPeopleintheFlowofMassCommunications.NewYork:TheFreePress.Leavitt,A.;Burchard,E.;Fisher,D.;andGilbert,S.2009.TheInuentials:NewApproachesforAnalyzingInuenceonTwitter.WebEcologyProject,http://tinyurl.com/lzjlzq.Leskovec,J.;Adamic,L.A.;andHuberman,B.A.2007.TheDynamicsofViralMarketing.ACMTrans.ontheWeb(TWEB).Levy,M.,andWindhal,S.1985.MediaGraticationsResearch,ChapterTheConceptofAudienceActivity.Sage.Rogers,E.M.1962.DiffusionofInnovations.FreePress.Watts,D.,andDodds,P.2007.Inuentials,Networks,andPublicOpinionFormation.JournalofConsumerResearch.Watts,D.2007.ChallengingtheInuentialsHypothesis.WordofMouthMarketingAssociation.Weng,J.;Lim,E.-P.;Jiang,J.;andHe,Q.2010.TwitterRank:FindingTopic-SensitiveInuentialTwitterers.InACMWSDM.