Earthquake Shakes Twitter Users Realtime Event Detection by Social Sensors Takeshi Sakaki - PDF document

EarthquakeShakesTwitterUsers:Real-timeEventDetectionbySocialSensorsTakeshiSakakiTheUniversityofTokyoYayoi2-11-16,Bunkyo-kuTokyo,Japantokyo.ac.jpMakotoOkazakiTheUniversityofTokyoYayoi2-11-16,Bunkyo-kuTokyo,Japan model.t.u-tokyo.ac.jpYutakaMatsuoTheUniversityofTokyoYayoi2-11-16,Bunkyo-kuTokyo,Japantokyo.ac.jpABSTRACT,apopularmicrobloggingservice,hasreceivedmuchattentionrecently.AnimportantcharacteristicofTwitter http://www.techcrunch.com/2009/08/03/twitter-reaches-44.5-million-people-worldwide-in-june-comscore/AccordingtoareportfromNielsen.com.www.tumblr.com,www.plurk.com,www.emote.in,www.squeelr.com,www.jaiku.com,identi.cahttp://mashable.com/2009/08/12/japan-earthquake/ JapanEarthquakeShakesTwitterUsers...AndBeyonce:EarthquakesareonethingyoucanbetonbeingcoveredonTwitter(Twitter)Þrst,because,quitefrankly,ifthegroundisshaking,youÕregoingtotweetaboutitbeforeitevenreg-isterswiththeUSGSandlongbeforeitgetsre-portedbythemedia.Thatseemstobethecaseagaintoday,asthethirdearthquakeinaweekhashitJapananditssurroundingislands,aboutanhourago.TheÞrstuserwecanÞndthattweetedaboutitwasRicardoDuranofScottsdale,AZ,who,judgingfromhisTwitterfeed,hasbeentrav-elingtheworld,arrivinginJapanyesterday.Thispostwellrepresentsthemotivationofourstudy.Theresearchquestionofourstudyis,Žcanwedetectsucheventoccurrenceinreal-timebymonitoringtweets?ŽThispaperpresentsaninvestigationofthereal-timena-tureofTwitterandproposesaneventnoti“cationsystemthatmonitorstweetsanddeliversnoti“cationpromptly.Toobtaintweetsonthetargeteventprecisely,weapplyse-manticanalysisofatweet:Forexample,usersmightmaketweetssuchasŽEarthquake!ŽorŽNowitisshakingŽthusearthquakecouldbekeywords,butusersmightalsomaketweetssuchasŽIamattendinganEarthquakeConferenceŽ,orŽSomeoneisshakinghandswithmybossŽ.Wepreparethetrainingdataanddeviseaclassi“erusingasupportvectormachinebasedonfeaturessuchaskeywordsinatweet,thenumberofwords,andthecontextoftarget-eventwords.Subsequently,wemakeaprobabilisticspatiotemporalmodelofanevent.Wemakeacrucialassumption:eachTwitteruserisregardedasaandeachtweetassensoryinfor-.Thesevirtualsensors,whichwecallsocialsensorsareofahugevarietyandhavevariouscharacteristics:somesensorsareveryactive;othersarenot.Asensorcouldbeinoperableormalfunctioningsometimes(e.g.,auserissleep-ing,orbusydoingsomething).Consequently,socialsensorsareverynoisycomparedtoordinalphysicalsensors.Regard-ingaTwitteruserasasensor,theeventdetectionproblemcanbereducedintotheobjectdetectionandlocationes-timationprobleminaubiquitous/pervasivecomputingen-vironmentinwhichwehavenumerouslocationsensors:auserhasamobiledeviceoranactivebadgeinanenviron-mentwheresensorsareplaced.Throughinfraredcommu-nicationoraWiFisignal,theuserlocationisestimatedasprovidinglocation-basedservicessuchasnavigationandmuseumguides[9,25].WeapplyKalman“ltersandparti-cle“lters,whicharewidelyusedforlocationestimationinubiquitous/pervasivecomputing.Asanapplication,wedevelopanearthquakereportingsystemusingJapanesetweets.BecauseofthenumerousearthquakesinJapanandthenumerousandgeographicallydispersedTwitterusersthroughoutthecountry,itissome-timespossibletodetectanearthquakebymonitoringtweets.Inotherwords,manyearthquakeeventsoccurinJapan.Manysensorsareallocatedthroughoutthecountry.Fig-ure1portraysamapofTwitterusersworldwide(obtainedfromUMBCeBiquityResearchGroup);Fig.2depictsamapofearthquakeoccurrencesworldwide(usingdatafromJapanMeteorologicalAgency(JMA)).Itisapparentthattheonlyintersectionofthetwomaps,whichmeansregionswithmanyearthquakesandlargeTwitterusers,isJapan.(OtherregionssuchasIndonesia,Turkey,Iran,Italy,andPaci“cUScitiessuchasLosAngelesandSanFranciscoalsoroughlyintersect,althoughthedensityismuchlowerthaninJapan.)Oursystemdetectsanearthquakeoccurrenceandsendsane-mail,possiblybeforeanearthquakeactuallyarrivesatacertainlocation:Anearthquakepropagatesatabout3…7km/s.Forthatreason,apersonwhois100kmdistantfromanearthquakehasabout20sbeforethearrivalofanearthquakewave.WepresentabriefoverviewofTwitterinJapan:TheJapaneseversionofTwitterwaslaunchedonApril2008.InFebruary2008,JapanwastheNo.2countrywithrespecttoTwittertrac.Atthetimeofthiswriting,Japanhasthe11thlargestnumberofusers(morethanhalfamillionusers)intheworld.Althougheventdetection(particularlytheearthquakedetection)iscurrentlypossiblebecauseofthehighdensityofTwitterusersandearthquakesinJapan,ourstudyisusefultodetecteventsofvarioustypesthroughouttheworld.Thecontributionsofthepaperaresummarizedasfollows:Thepaperprovidesanexampleofintegrationofse-manticanalysisandreal-timenatureofTwitter,andpresentspotentialusesforTwitterdata.Forearthquakepredictionandearlywarning,manystudieshavebeenmadeintheseismology“eld.Thispaperpresentsaninnovativesocialapproach,whichhasnotbeenreportedbeforeintheliterature.Thispaperisorganizedasfollows:Inthenextsection,weexplainsemanticanalysisandsensoryinformation,followedbythespatiotemporalmodelinSection3.InSection4,wedescribetheexperimentsandevaluationofeventdetection.TheearthquakereportingsystemisintroducedintoSection5.Section6isdevotedtorelatedworksanddiscussion.Finally,weconcludethepaper.2.EVENTDETECTIONInthispaper,wetargeteventdetection.Anisanar-bitraryclassi“cationofaspace/timeregion.Aneventmighthaveactivelyparticipatingagents,passivefactors,products,andalocationinspace/time[21].Wetargeteventssuchasearthquakes,typhoons,andtracjams,whicharevisiblethroughtweets.Theseeventshaveseveralproperties:i)theyareoflargescale(manyusersexperiencetheevent),ii)theyparticularlyin”uencepeoplesdailylife(forthatreason,theyareinducedtotweetaboutit),andiii)theyhavebothspatialandtemporalregions(sothatreal-timelocationestimationwouldbepossible).Sucheventsincludesocialeventssuchaslargeparties,sportsevents,exhibi-tions,accidents,andpoliticalcampaigns.Theyalsoincludenaturaleventssuchasstorms,heavyrainfall,tornadoes,typhoons/hurricanes/cyclones,andearthquakes.Wedes-ignateaneventwewouldliketodetectusingTwitterasatargetevent2.1SemanticAnalysisonTweetTodetectatargeteventfromTwitter,wesearchfromTwitterand“ndusefultweets.Tweetsmightincludemen-tionsofthetargetevent.Forexample,usersmightmaketweetssuchasŽEarthquake!ŽorŽNowitisshakingŽ.Con-sequently,earthquakecouldbekeywords(whichwecallquerywords).butusersmightalsomaketweetssuchasŽIamattendinganEarthquakeConferenceŽ,orŽSome-oneisshakinghandswithmybossŽ.Moreover,evenifa http://blog.twitter.com/2008/02/twitter-web-trac-around-world.html Figure1:Twitterusermap. Figure2:Earthquakemap.tweetisreferringtothetargetevent,itmightnotbeappro-priateasaneventreport;forexampleausermakestweetssuchasŽTheearthquakeyesterdaywasscaringŽ,orŽThreeearthquakesinfourdays.Japanscaresme.ŽThesetweetsaretrulythementionsofthetargetevent,buttheyarenotreal-timereportsoftheevents.Therefore,itisnecessarytoclarifythatatweetisactuallyreferringtoanactualearth-quakeoccurrence,whichisdenotedasapositiveclass.Toclassifyatweetintoapositiveclassoranegativeclass,weuseasupportvectormachine(SVM)[14],whichisawidelyusedmachine-learningalgorithm.Bypreparingpos-itiveandnegativeexamplesasatrainingset,wecanpro-duceamodeltoclassifytweetsautomaticallyintopositiveandnegativecategories.Wepreparethreegroupsoffeaturesforeachtweetasfol-lows:FeaturesA(statisticalfeatures)thenumberofwordsinatweetmessage,andthepositionofthequerywordwithinatweet.FeaturesB(keywordfeatures)thewordsinatweetFeaturesC(wordcontextfeatures)thewordsbeforeandafterthequeryword.TohandleJapanesetexts,morphologicalanalysisiscon-ductedusingMecab,whichseparatessentencesintoasetofwords.InthecaseofEnglish,weapplyastandardstop-wordeliminationandstemming.WecomparetheusefulnessofthefeaturesinSection4.Usingtheobtainedmodel,wecanclassifywhetheranewtweetcorrespondstoapositiveclassoranegativeclass. Becauseatweetisusuallyshort,weuseeverywordinatweetbyconvertingitintoawordID.http://mecab.sourceforge.net/2.2TweetasaSensoryValueWecansearchthetweetandclassifyitintoapositiveclassifausermakesatweetonatargetevent.Inotherwords,theuserfunctionsasaoftheevent.Ifshemakesatweetaboutanearthquakeoccurrence,thenitcanbeconsideredthatshe,asanŽearthquakesensorŽ,returnsapositivevalue.AtweetcanthereforebeconsideredasasensorreadingThisisacrucialassumption,butitenablesapplicationofvariousmethodsrelatedtosensoryinformation.Assumption2.1EachTwitteruserisregardedasasen-sor.Asensordetectsatargeteventandmakesareportprobabilistically.Thevirtualsensors(orsocialsensors)havevariouschar-acteristics:somesensorsareactivated(i.e.maketweets)onlyaboutspeci“cevents,althoughothersareactivatedtoawiderrangeofevents.Thenumberofsensorsislarge;therearemorethan40millionsensorsworldwide.Asen-sormightbeinoperableoroperatingincorrectlysometimes(whichmeansauserisnotonline,sleeping,orisbusydo-ingsomething).Therefore,thissocialsensorisnoisierthanordinalphysicalsensorssuchaslocationsensors,thermalsensors,andmotionsensors.Atweetcanbeassociatedwithatimeandlocation:eachtweethasitsposttime,whichisobtainableusingasearchAPI.Infact,GPSdataareattachedtoatweetsometimes,e.g.whenauserisusinganiPhone.Alternatively,eachTwitterusermakesaregistrationontheirlocationintheuserpro“le.Theregisteredlocationmightnotbethecurrentlocationofatweet;however,wethinkitisprobablethatapersonisneartheregisteredlocation.Inthisstudy,weuseGPSdataandtheregisteredlocationofauser.Wedonotusethetweetforspatialanalysisifthelocationisnotavailable(WeusethetweetinformationfortemporalAssumption2.2Eachtweetisassociatedwithatimeandlocation,whichisasetoflatitudeandlongitude.Byregardingatweetasasensoryvalueassociatedwithalocationinformation,theeventdetectionproblemisre-ducedtodetectinganobjectanditslocationfromsensorreadings.Estimatinganobjectslocationisarguablythemostfundamentalsensingtaskinmanyubiquitousandper-vasivecomputingscenarios[7].Figure3presentsanillustrationofthecorrespondencebetweensensorydatadetectionandtweetprocessing.Themotivationsarethesameforbothcases:todetectatargetevent.Observationbysensorscorrespondstoanobserva-tionbyTwitterusers.Theyareconvertedintovaluesbyaclassi“er.Aprobabilisticmodelisusedtodetectanevent,asdescribedinthenextsection.3.MODELInorderforeventdetectionandlocationestimation,weuseprobabilisticmodels.Inthissection,we“rstdescribeeventdetectionfromtime-seriesdata.Then,wedescribethelocationestimationofatargetevent.3.1TemporalModelEachtweethasitsposttime.Whenatargeteventoc-curs,howcanthesensorsdetecttheevent?Wedescribethetemporalmodelofeventdetection.First,weexaminetheactualdata.Figures4and5re-spectivelypresentthenumberstweetsfortwotargetevents: Figure3:CorrespondencebetweeneventdetectionfromTwitterandobjectdetectioninaubiquitousenvironment.anearthquakeandatyphoon.Itisapparentthatspikesoccuronthenumberoftweets.Eachcorrespondstoaneventoccurrence.Inthecaseofanearthquake,morethan10earthquakesoccurduringtheperiod.Inthecaseofty-phoon,Japansmainpopulationcenterswerehitbyalargetyphoon(designatedasMelor)inOctober2009.Thedistributionisapparentlyanexponentialdistribu-tion.Theprobabilitydensityfunctionoftheexponentialdistributionis0andTheexponentialdistributionoccursnaturallywhendescrib-ingthelengthsoftheinter-arrivaltimesinahomogeneousPoissonprocess.IntheTwittercase,wecaninferthatifauserdetectsaneventattime0,assumethattheprobabilityofhispostingatweetfromtois“xedas.Then,thetimetomakeatweetcanbeconsideredasanexponentialdistribution.Evenifauserdetectsanevent,therefore,shemightnotmakeatweetrightawayifsheisnotonlineordoingsome-thing.Shemightmakeapostonlyaftersuchproblemsareresolved.Therefore,itisreasonablethatthedistributionofthenumberoftweetsfollowsanexponentialdistribution.Actuallythedata“tsverywelltoanexponentialdistribu-tion;weget34withonaverageToassessanalarm,wemustcalculatethereliabilityofmultiplesensorvalues.Forexample,ausermightmakeafalsealarmbywritingatweet.Itisalsopossiblethattheclassi“ermisclassi“esatweetintoapositiveclass.WecandesignthealarmprobabilisticallyusingthefollowingtwoThefalse-positiveratioofasensorisapproximately0.35,asweshowinSection4.1.Sensorsareassumedtobeindependentandidenticallydistributed(i.i.d.),asweexplaininSection3.3.Assumingthatwehavesensors,whichproducepositivesignals,theprobabilityofallsensorsreturningafalse- Figure4:Numberoftweetsrelatedtoearthquakes. Figure5:Numberoftweetsrelatedtotyphoons.alarmis.Therefore,theprobabilityofeventoccurrencecanbeestimatedas1.Givensensorsattime0sensorsattime.Therefore,thenumberofsensorsweexpectattimeConsequently,theprobabilityofaneventoccurrenceattimeWecancalculatetheprobabilityofeventoccurrenceifwe34and35.Forexample,ifwereceivepositivetweetsandwouldliketomakeanalarmwithafalse-positiveratiolessthan1%,wecancalculatetheexpectedwaittimewaittodeliverthenoti“cationaswait=(1Althoughmanyworksdescribingeventdetectionhavebeenreportedinthedatamining“eld,weusethissimpleap-proachutilizingthecharacteristicsoftheclassi“erandthe3.2SpatialModelEachtweetisassociatedwithalocation.Wedescribehowtoestimatethelocationofaneventfromsensorreadings.Tode“netheproblemoflocationestimation,weconsidertheevolutionofthestatesequenceofatarget,givenisapossiblynonlinearfunctionofthestate.Furthermore,isani.i.dprocessnoisesequence.Theobjectiveoftrackingistoestimaterecursivelyfrommeasurements),whereisapossiblynonlinearfunction,andwhereisani.i.dmeasurementnoisesequence.FromaBayesianperspective,thetrackingproblemistocalculaterecursivelysomedegreeofbeliefinthestateattime,givendatauptotime Presumingthat)isavailable,thepredictionstageusesthefollowingequation:.HereweuseaMarkovprocessoforderone.Therefore,wecanassumeInupdatestage,theBayesruleisappliedaswherethenormalizingconstantTosolvetheproblem,severalmethodsofBayesian“ltersareproposedsuchasKalman“lters,multi-hypothesistrack-ing,grid-basedandtopologicalapproaches,andparticle“l-ters[7].Forthisstudy,weuseKalman“ltersandparticle“lters,bothofwhicharewidelyusedinlocationestimation.3.2.1KalmanFiltersTheKalman“lterassumesthattheposteriordensityateverytimestepisGaussianandthatitisthereforeparam-eterizedbyameanandcovariance.Wecanwriteitas.Therein,areknownmatricesde“ningthelinearfunctions.Thecovariantsofare,respectively,TheKalman“lteralgorithmcanconsequentlybeviewedasthefollowingrecursiverelation:),and,andwherem,P)isaGaussiandensitywithargument,mean,covariance,andforwhichthefollowingaretrue:,andThisistheoptimalsolutiontothetrackingproblemiftheassumptionshold.AKalman“lterworksbetterinalinearGaussianenvironment.WhenutilizingKalman“lters,itisimportanttoconstructagoodmodelandparameters.Inthispaper,weimplementmodelsfortwocasesasfollows.Case1:Locationestimationofanearthquakecenter.Inthiscase,weneednottakeintoconsiderationthetime-transitionproperty,thusweuseonlylocationinformationWesetisthelongitudeisthelatitude;,and=0.Weassumethaterrorsoftemporaltransitiondonotoccur,anderrorsinobservationareGaussianforsimplicity:=0,,2],andCase2:Trajectoryestimationofatyphoon.Weneedtoconsiderboththelocationandthevelocityofanevent.WeapplytheNewtonsmotionequationasfollows:isthevelocityonlongitude,isthevelocityonlatitude.Weset01000100001 2t2,ayt t,aistheaccelera-tiononlongitude,andistheaccelerationonlatitude.SimilarlyasinCase1,weassumethaterrorsoftemporaltransitiondonotoccurr,anderrorsinobservationareGaus-sianforsimplicity:=0,,2],and3.2.2ParticleFiltersAparticle“lterisaprobabilisticapproximationalgorithm Algorithm1Particle“lteralgorithm :Calculatetheweightdistributionx,yfromtwitterusersgeographicdistributioninJapan.:Generateandweightaparticleset,whichdiscretehypothesis.(1)Generateaparticleset)andallocatethemonthemapevenly:particle,weightcorrespondstothelongitudeandspondstothelatitude.(2)Weightthembasedonweightdistributionx,y(1)Re-sampleparticlesfromaparticlesetweightsofeachparticlesandallocatethemonthemap.(Weallowtore-samplesameparticlesmorethan(2)Generateanewparticlesetandweightthembasedonweightdistributionx,y:PredictthenextstateofaparticlesettheNewton’smotionequation. y,ty,t y,ty,ty,ty,tWeighing:Re-calculatetheweightofbymeasurement)asfollows.,dy ( 2·Š(k+k) Measurement:Calculatethecurrentobjectlocation)bytheaverageof:IterateStep3,4,5and6untilconvergence. implementingaBayes“lter,andamemberofthefamilyofsequentialMonteCarlomethods.Forlocationestima-tion,itmaintainsaprobabilitydistributionfortheloca-tionestimationattime,designatedasthebeliefBel.Eachisadiscretehypothesisaboutthelocationoftheobject.Thearenon-negativeweights,importancefactors,whichsumtoone.TheSequentialImportanceSampling(SIS)algorithmisaMonteCarlomethodthatformsthebasisforparticle“lters.TheSISalgorithmconsistsofrecursivepropagationoftheweightsandsupportpointsaseachmeasurementisreceivedsequentially.Weuseamoreadvancedalgorithmwithre-sampling[1].Weemployweightdistributionx,y)whichisobtainedfromtwitteruserdistributiontotakeintocon-siderationthebiasesofuserlocationsThealogorithmisshowninAlgo.1.3.3InformationDiffusionrelatedtoaReal-timeEventSomeinformationrelatedtoaneventdiusesthroughTwitter.Forexample,ifauserdetectsanearthquakeand Wesampletweetsassociatedwithlocationsandgetuserdistributionproportionaltothenumberoftweetsineach Figure6:Earthquakeinforma-tiondiusionnetwork. Figure7:Typhooninformationdiusionnetwork. Figure8:AnewNintendogameinformationdiusionnetwork.makesatweetabouttheearthquake,afollowerofthatusermightmaketweetsaboutthat.Thischaracteristicisimpor-tantbecause,inourmodel,sensorsmightnotbeindepen-denteachother,whichwouldcauseanundesirableeectoneventdetection.Figures6,7,and8respectivelyportraytheinformation”ownetworkonearthquake,typhoon,andanewNintendoDSgame.Weinferthenetworkasfollows:AssumethatuserAfollowsuserB.IfuserBmakesatweetaboutanevent,andsoonafterthatifuserAmakesatweetaboutanevent,thenweconsidertheinformation”owsfromBtoAThisisthesimilarde“nitiontootherstudiesofinformationdiusion(e.g.,[15,16]).Wecanunderstandthat,inthecaseofearthquakesandtyphoons,verylittleinformationdiusiontakesplaceonTwitter.Ontheotherhand,thereleaseofanewgameillustratesthescaleandrapidityofinformationdiusion.Therefore,wecanassumethatthesensorsarei.i.d.whenconsideringreal-timeeventdetectionsuchastyphoonsandearthquakes.4.EXPERIMENTSANDEVALUATIONInthissection,wedescribetheexperimentalresultsandevaluationoftweetclassi“cationandlocationestimation.ThewholealgorithmisshowninAlgo.2.Weprepareasetofqueriesforantargetevent.We“rstsearchfortweetsincludingthequerysetfromTwittereveryWeuseasearchAPItosearchtweets.Intheearthquakecase,wesetŽearthquakeŽandŽshakingŽandinthetyphooncase,wesetŽtyphoonŽ.Wesetas3s.Afterdeterminingaclassi“cationandobtainingapositiveexample,thesystemmakesacalculationofatemporalandspatialprobabilisticmodel.Weconsiderthataneventisdetectediftheprobabilityishigherthanacertainthreshold95inourcase).Thelocationinformationofeachtweetisobtainedandusedforlocationestimationoftheevent.Intheearthquakereportingsystemexplainedinthenextsection,thesystemquicklysendsane-mail(usuallymobilee-mail)toregisteredusers.4.1EvaluationbySemanticAnalysis LovePlus,agamethatoersavirtualgirlfriendexperience,whichwasrecentlyreleasedinSeptember3,2009.Becauseofthisde“nition,thediusionincludesretweetwhichisatypeofmessagethatrepeatssomeinformationthatwaspreviouslytweetedbyanotheruser.search.twitter.com Algorithm2Eventdetectionandlocationestimational- 1.Givenasetofqueriesforatargetevent.2.PutaqueryusingsearchAPIeverysecondsandobtaintweets3.Foreachtweet,obtainfeatures,and.Applytheclassi“cationtoobtainvalue4.Calculateeventoccurrenceprobability;ifitisabovethethresholdthre,thenproceedtostep5.Foreachtweet,weobtainthelatitudeandthelon-byi)utilizingtheassociatedGPSlocation,ii)makingaquerytoGoogleMaptheregisteredlocationfor.Set=nullifbothdonotwork.6.CalculatetheestimatedlocationoftheeventfromusingKalman“lteringorparticle“ltering.7.(optionally)Sendalerte-mailstoregisteredusers. Forclassi“cationoftweets,weprepared597positiveex-ampleswhichreportearthquakeoccurrenceasatrainingset.Theclassi“cationperformanceispresentedinTable1.Weusetwoquerywords„earthquake;performancesusingeitherqueryareshown.WeusedalinearkernelforSVM.Weobtainthehighest-valuewhenweusefeatureAandallfeatures.Surprisingly,featureBandfeatureCdonotcontributemuchtotheclassi“cationperformance.Whenanearthquakeoccurs,auserbecomessurprisedandmightproduceaveryshorttweet.Itisapparentthattherecallisnotsohighasprecision.Itisattributabletotheusageofquerywordsinadierentcontextthanweintend.Sometimesitisdicultevenforhumanstojudgewhetheratweetisreportinganactualearthquakeornot.Someex-amplesarethatausermightwriteŽIsthisanearthquakeoratruckpassing?ŽOverall,theclassi“cationperformanceisgoodconsideringthatwecanusemultiplesensorreadingsasevidenceforeventdetection.4.2EvaluationofSpatialEstimationFigure9presentsthelocationestimationofanearthquakeonAugust11.Wecan“ndthatmanytweetsoriginatefromawideregioninJapan.Theestimatedlocationoftheearth-quake(shownasestimationbyparticle“lter)isclosetotheactualcenteroftheearthquake,whichshowstheeciencyofthelocationestimationalgorithm.Table2presentsre- Wedonotshowtheresultforthetyphooncasebecauseofspacelimitations. Table1:Performanceofclassi“cation.earthquakeFeaturesRecallPrecision-value A87.50%63.64%73.69%B87.50%38.89%53.85%C50.00%66.67%57.14%All87.50%63.64%73.69% FeaturesRecallPrecision-value A66.67%68.57%67.61%B86.11%57.41%68.89%C52.78%86.36%68.20%All80.56%65.91%72.50% Figure9:Earthquakelocationestimationbasedontweets.Balloonsshowthetweetsontheearthquake.Thecrossshowstheearthquakecenter.Redrepre-sentsearlytweets;bluerepresentslatertweets.sultsoflocationestimationfor25earthquakesinAugust,September,andOctober2009.WecompareKalman“lter-ingandparticle“ltering,withtheweightedaverageandthemedianasabaseline.Theweightedaveragesimplytakestheaverageoflatitudesandlongitudeonallthepositivetweets,andmediansimplytakesthemedianofthem.Particle“ltersperformwellcomparedtoothermethods.Thepoorperfor-manceofKalman“lteringimpliesthatthelinearGaussianassumptiondoesnotholdforthisproblem.Wecan“ndthatifthecenteroftheearthquakeisintheseaarea,itismorediculttolocateitpreciselyfromtweets.Similarly,itbecomesmorediculttomakegoodestimationsinless-populatedareas.Thatisreasonable:allotherthingsbeingequal,thegreaterthenumberofsensors,themoreprecisetheestimationwillbe.Figure10isthetrajectoryestimationoftyphoonMelorbasedontweets.Inthecaseofanearthquake,thecenterisonelocation.However,inthecaseofatyphoon,thecentermovesandmakesatrajectory.ThecomparisonoftheperformanceisshowninTable3.Theparticle“lterworkswellandoutputsasimilartrajectorytotheactualtrajectory.5.EARTHQUAKEREPORTINGSYSTEMWedevelopedanearthquakereportingsystemusingtheeventdetectionalgorithm.Earthquakeinformationismuch Figure10:Typhoontrajectoryestimationbasedontweets.morevaluableifgiveninrealtime.Wecanturnoastoveorheaterinourhouseandhideourselvesunderadeskortableifwehaveseveralsecondsbeforeanearthquakeactu-allyhits.SeveralTwitteraccountsreportearthquakeoccur-rence.SomeexamplesarethattheUnitedStatesGeologicalSurvey(USGS)feedstweetsonworldearthquakeinforma-tion,butitisnotusefulforpredictionorearlywarning.Vastamountsofworkhavebeendoneonintermediate-termearthquakepredictionintheseismology“eld(e.g.[23]).Variousattemptshavealsobeenmadetoproduceshort-termforecaststorealizeanearthquakewarningsystembyobservingelectromagneticemissionsfromground-basedsen-sorsandsatellites[3].Otherprecursorsignalssuchasiono-sphericchanges,infraredluminescence,andair-conductivitychange,alongwithtraditionalmonitoringofmovementsoftheearthscrust,areinvestigated.InJapan,thegovernmenthasallocatedaconsiderableamountofitsbudgettomitigatingearthquakedamage.AnearthquakeearlywarningservicehasbeenoperatedbyJMAsince2007.Itprovidesadvanceannouncementsofthees-timatedseismicintensitiesandexpectedarrivaltimes.ItdetectsP-waves(primarywaves)andmakesanalertimme-diatelysothatearthquakedamagecanbemitigatedthroughcountermeasuressuchasslowingtrainsandcontrollingel-evators.Infact,P-wavesareatypeofelasticwavethatcantravelfasterthantheS-waves(secondarywaves),whichcausesheareectsandengendermuchmoredamage.Theproposedsystem,calledToretter,hasbeenoperatedsinceAugust8ofthisyear.AsystemscreenshotisdepictedinFig.11.Userscanseethedetectionofpastearthquakes.Theycanregistertheire-mailstoreceivenoticesoffutureearthquakedetectionreports.Asamplee-mailispresentedinFig.12.Italertsusersandurgesthemtopreparefortheearthquake.Itishopedthatthee-mailisreceivedbyausershortlybeforetheearthquakeactuallyarrives.Anearthquakeistransmittedthroughtheearthscrustatabout3…7km/s.Therefore,apersonhasabout20sbeforeitsarrivalatapointthatis100kmdistant.Table4presentssomefactsaboutearthquakedetectionandnoti“cationusingoursystem.Thistableshowsthatweinvestigated10earthquakesduring18August…2Septem-ber,allofwhichoursystemdetected.The“rsttweetof ItmeansŽwehavetakenitŽinJapanese. Table2:Locationestimationaccuracyofearthquakesfromtweets.Foreachmethod,weshowthedierenceoftheestimatedlatitudeandthelongitudetotheactualones,andtheEucliddistanceofthem.Smallerdistancemeansbetterperformance. Actualcenter Median(baseline) Weightedave.(baseline) Kalman“lters ParticleÞlters lat.long. lat.long.dist. lat.long.dist. lat.long.dist. lat.long.dist. Aug.1001:00 33.10138.50 3.40-0.803.49 2.70-0.102.70 2.67-0.502.72 2.600.50Aug.1105:00 34.80138.50 0.90-0.901.27 0.70-0.300.76 0.60-0.20 0.30-0.900.95Aug.1307:50 33.00140.80 1.30-9.609.69 2.30-2.30 1.63-3.754.09 2.70-2.703.82Aug.1720:40 33.70130.20 4.606.007.56 0.903.203.32 1.634.354.65 0.10-0.80Aug.1822:17 23.30123.50 7.809.9012.60 8.7010.9013.95 8.3210.1313.11 5.608.10Aug.2108.51 35.70140.00 0.50-4.404.43 0.10-1.001.00 0.00-0.60 -0.800.480.93Aug.2413:30 37.50138.60 -0.400.00 -0.500.400.64 -0.500.300.58 2.400.702.50Aug.2414:40 41.10140.30 -1.901.102.20 -1.300.50 -1.500.501.58 3.102.003.69Aug.2502:22 42.10142.80 -2.90-3.904.86 -6.10-3.807.19 -5.20-3.706.38 -1.80-1.90Aug.2520:19 35.40140.40 1.60-1.802.41 2.20-0.702.31 0.70-1.601.75 1.400.10Aug.3100:46 37.20141.50 -0.40-3.603.62 -1.10-2.302.55 -1.30-2.202.56 -0.30-0.30Aug.3121:11 33.40130.90 -4.50-3.605.76 0.502.102.16 0.701.902.02 -0.20-1.70Sep.322:26 31.10130.30 6.20-0.106.20 4.005.006.40 4.907.208.71 2.402.10Sep.411:30 35.80140.10 3.10-1.703.54 0.20-0.90 0.00-1.001.00 0.801.401.61Sep.0510:59 37.00140.20 -2.70-8.308.73 -1.40-3.10 -1.30-3.303.55 -2.10-5.806.17Sep.0801:24 42.20143.00 -3.60-8.909.60 -2.50-3.904.63 -4.50-6.007.50 1.30-3.60Sep.1018:29 43.20146.20 -5.90-10.2011.78 -4.90-7.108.63 -4.50-7.208.49 -0.90-7.00Sep.1621:38 33.40130.90 1.10-0.20 0.902.102.28 0.501.401.49 -0.20-2.502.51Sep.2220:40 47.60141.70 -11.10-7.5013.40 -10.80-3.1011.24 -11.30-3.8011.92 -7.80-3.00Oct.119:43 36.40140.70 0.70-3.803.86 -0.60-1.801.90 -0.30-1.501.53 -0.700.30Oct.509:35 42.40141.60 -3.70-3.104.83 -2.70-2.003.36 -2.60-1.603.05 1.10-1.70Oct.607:49 35.90137.60 0.501.201.30 -0.200.800.82 -0.100.900.91 0.300.50Oct.1017:43 41.80142.20 -3.50-5.406.44 -1.40-2.10 -2.20-2.603.41 2.40-1.302.73Oct.1216:10 35.90137.60 2.800.502.84 0.801.20 0.801.601.79 3.601.403.86Oct.1218:42 37.40139.70 -2.00-4.404.83 -1.50-0.901.75 -1.70-1.402.20 -1.00-0.60 Averagedistance 3.62 3.85 Table3:TrajectoryestimationaccuracyoftyphoonMelorfromtweets. Location Median(baseline) Weightedave.(baseline) Kalman“lters ParticleÞlters lat.long. lat.long.dist. lat.long.dist. lat.long.dist. lat.long.dist. Oct.712:00 29.00131.80 -1.90-1.90 -5.20-3.606.32 -3.90-1.104.05 -4.701.104.83Oct.715:00 29.90132.50 -3.70-2.604.52 -3.80-2.404.49 3.203.104.46 -2.700.90Oct.718:00 30.80133.20 -4.10-1.904.52 -4.40-3.505.62 -6.405.408.37 -3.20-0.70Oct.721:00 31.60134.30 -3.90-3.505.24 -3.60-3.304.88 -10.90-1.6011.02 -3.70-0.50Oct.80:00 32.90135.60 -2.30-0.10 -2.30-0.902.47 -12.60-20.4023.98 -2.90-3.504.55Oct.86:00 35.10137.20 1.603.003.40 0.801.70 4.2016.0016.54 -0.60-2.502.57Oct.89:00 36.10138.80 -0.603.603.65 0.000.50 0.502.602.65 0.70-0.801.06Oct.812:00 37.10139.70 1.703.904.25 1.501.201.92 2.101.602.64 1.400.10Oct.815:00 38.00140.90 2.303.203.94 2.402.20 1.707.607.79 2.402.703.61Oct.818:00 39.00142.30 3.207.307.97 3.505.10 2.10-18.8018.92 3.705.106.30Oct.821:00 40.00143.60 4.303.905.81 4.005.306.64 1.604.504.78 4.203.10 Averagedistance 4.02 9.56 Table5:EarthquakedetectionperformancefortwomonthsfromAugust2009.JMAintensityscale 2ormore3ormore4ormore Num.ofearthquakes 78253 70(89.7%)24(96.0%)3(100.0%)Promptlydetected 53(67.9%)20(80.0%)3(100.0%) anearthquakeisusuallymadewithinaminuteorso.Thedelaycanresultfromthetimeforpostingatweetbyauser,thetimetoindexthepostinTwitterservers,andthetimetomakequeriesbyoursystem.Weapplyclassi“cationfor49,314tweetsretrievedbyquerywordsinonemonth;re-sultsshow6,291positivetweetspostedby4,218users.Ev-eryearthquakeelicitedmorethan10tweetswithin10min,exceptoneinBungo-suido,whichistheseabetweentwolargeislands:KyushuandShikoku.Oursystemsente-mailsmostlywithinaminute,sometimeswithin20s.ThedeliverytimeisfarfasterthantherapidbroadcastofannouncementofJMA,whicharewidelybroadcastonTV;onaverage,aJMAannouncementisbroadcast6minafteranearthquakeoccurs.Statistically,wedetected96%ofearthquakeslargerthanJMAseismicintensityscale3ormoreasshowninTable5.6.RELATEDWORKTwitterisaninterestingexampleofthemostrecentsocialmedia:numerousstudieshaveinvestigatedTwitter.AsidefromthestudiesintroducedinSection1,severalothershavebeendone.Grossecketal.investigatedindicatorssuchasthein”uenceandtrustrelatedtoTwitter[8].Krish-namurthyetal.crawlednearly100,000Twitterusersandexaminedthenumberofuserseachuserfollows,inaddi-tiontothenumberofusersfollowingthem.Naamanetal.analyzedcontentsofmessagesfrommorethan350Twitter TheJMAseismicintensityscaleisameasureusedinJapanandTaiwantoindicateearthquakestrength.UnliketheRichtermagnitudescale,theJMAscaledescribesthedegreeofshakingatapointontheearthssurface.Forexample,theJMAscale3is,byde“nition,onewhichisŽfeltbymostpeopleinthebuilding.SomepeoplearefrightenedŽ.ItissimilartotheModi“edMercalliscaleIV,whichisusedalongwiththeRichterscaleintheUS. Table4:Factsaboutearthquakedetection.DateMagnitudeLocationTimeE-mailsenttime#tweetswithin10minAnnounceofJMA Aug.184.5Tochigi6:58:557:00:303507:08Aug.183.1Suruga-wan19:22:4819:23:141719:28Aug.214.1Chiba8:51:168:51:35528:56Aug.254.3Uraga-oki2:22:492:23:212302:27Aug.253.5Fukushima22:21:1622:22:291322:26Aug.273.9Wakayama17:47:3017:48:111617:53Aug.272.8Suruga-wan20:26:2320:26:451420:31Aug.314.5Fukushima00:45:5400:46:243200:51Sep.23.3Suruga-wan13:04:4513:05:041813:10Sep.23.6Bungo-suido17:37:5317:38:27317:43 Figure11:ScreenshotofToretter,anearthquakereportingsystem.  DearAlice,WehavejustdetectedanearthquakearoundChiba.Pleasetakecare.ToretterAlertSystem  Figure12:Samplealerte-mail.usersandmanuallyclassi“edmessagesintoninecategories[19].ThenumerouscategoriesareŽMenowŽandŽState-mentsandRandomThoughtsŽ;statementsaboutcurrenteventscorrespondingtothiscategory.SomestudiesattempttoshowapplicationsofTwitter:Borauetal.triedtouseTwittertoteachEnglishtoEnglish-languagelearners[4].Ebneretal.investigatedtheap-plicabilityofTwitterforeducationalpurposes,i.e.mobilelearning[6].TheintegrationoftheSemanticWebandmi-crobloggingwasdescribedinapreviousstudy[20]inwhichadistributedarchitectureisproposedandthecontentsareaggregated.Jensenetal.analyzedmorethan150thousandtweets,particularlythosementioningbrandsincorporateaccounts[12].IncontrasttothesmallnumberofacademicstudiesofTwitter,manyTwitterapplicationsexist.SomeareusedforanalysesofTwitterdata.Forexample,Tweettronicsprovidesananalysisoftweetsrelatedtobrandsandprod-uctsformarketingpurposes.Itcanclassifypositiveandnegativetweets,andcanidentifyin”uentialusers.Theclas- http://www.tweettronics.comsi“cationoftweetsmightbedonesimilarlytoouralgorithm.Web2expressDigestisawebsitethatauto-discoversinfor-mationfromTwitterstreamingdatato“ndreal-timeinter-estingconversations.Italsousesnaturallanguageprocess-ingandsentimentanalysistodiscoverinterestingtopics,aswedoinourstudy.Variousstudieshavebeenmadeoftheanalysisofwebdata(exceptforTwitter)particularlyaddressingthespatialaspect:ThemostrelevantstudytooursisonebyBack-strometal.[2].Theyusequerieswithlocation(obtainedbyIPaddresses),anddevelopaprobabilisticframeworkforquantifyingspatialvariation.Themodelisbasedonade-compositionofthesurfaceoftheearthintosmallgridcells;theyassumethatforeachgridcell,thereisaprobabil-itythatarandomsearchfromthiscellwillbeequaltothequeryunderconsideration.Theframework“ndsaquerysgeographiccenterandspatialdispersion.Exam-plesincludebaseballteams,newspapers,universities,andtyphoons.Althoughthemotivationisverysimilar,eventstobedetecteddier.Someexamplesarethatpeoplemightnotmakeasearchqueryearthquakewhentheyexperienceanearthquake.Therefore,ourapproachcomplementstheirwork.Similarlytoourwork,Meietal.targetedblogsandanalyzedtheirspatiotemporalpatterns[17].TheypresentedexamplesforHurricaneKatrina,HurricaneRita,andiPodNano.Themotivationofthatstudyissimilartoours,butTwitterdataaremoretime-sensitive;ourstudyexaminesevenmoretime-criticaleventse.g.earthquakes.Someworkshavetargetedcollaborativebookmarkingdata,asFlickrdoes,fromaspatiotemporalperspective:Serdyukovetal.investigatedgenericmethodsforplacingphotographsonFlickrontheworldmap[24].Theyusedalanguagemodeltoplacephotos,andshowedthattheycaneectivelyestimatethelanguagemodelthroughanalysesofannota-tionsbyusers.Rattenburyetal.[22]speci“callyexaminedtheproblemofextractingplaceandeventsemanticsfortagsthatareassignedtophotographsonFlickr.Theyproposedscale-structureidenti“cation,whichisaburst-detectionmethodbasedonscaledspatialandtemporalsegments.Locationestimationstudiesareoftendoneinthe“eldofubiquitouscomputing.Estimatinganobjectslocationisarguablythemostfundamentalsensingtaskinmanyubiq-uitousandpervasivecomputingscenarios.Representinglo-cationsstatisticallyenablesauni“edinterfaceforlocationinformation,whichenablesustomakeapplicationsindepen-dentofthesensorsused„evenwhenusingverydierentsensortypes,suchasGPSandinfraredbadges[7],orevenTwitter.WellknownalgorithmsforlocationestimationareKalman“lters,multihypothesistracking,grid-based,andtopologicalapproaches,andparticle“lters.HightowerandBorriellomadeastudyofapplyingparticle“lterstolocationsensorsdeployedthroughoutalabbuilding[10].Morethan http://web2express.org 30labresidentsweretracked;theirlocationswereestimatedaccuratelyusingtheparticle“lterapproach.7.DISCUSSIONWeplantoexpandoursystemtodetecteventsofvariouskindsusingTwitter.Wedevelopedanotherprototypethatdetectsrainbowinformation.Arainbowmightbevisiblesomewhereintheworld;someonemightbetwitteringaboutarainbow.Oursystemcanidentifyrainbowtweetsusingasimilarapproachtothatusedfordetectingearthquakes.Thedierencesarethatintherainbowcase,theinformationisnotsotime-sensitiveasthatintheearthquakecase.Ourmodelincludestheassumptionthatasingleinstanceofthetargeteventexists.Forexample,weassumethatwedonothavetwoormoreearthquakesortyphoonssimulta-neously.Althoughtheassumptionisreasonableforthesecases,itmightnotholdforothereventssuchastracjams,accidents,andrainbows.Torealizemultipleeventdetec-tion,wemustproduceadvancedprobabilisticmodelsthatallowhypothesesofmultipleeventoccurrences.Asearchqueryisimportanttosearchpossibly-relevanttweets.Forexample,wesetaquerytermasearthquakebecausemosttweetsmentioninganearthquakeoccurrenceuseeitherword.However,toimprovetherecall,itisnecessarytoobtainagoodsetofqueries.Wecanuseadvancedalgorithmsforqueryexpansion,whichisasubjectofourfuturework.8.CONCLUSIONAsdescribedinthispaper,weinvestigatedthereal-timenatureofTwitter,inparticularforeventdetection.Seman-ticanalyseswereappliedtotweetstoclassifythemintoapositiveandanegativeclass.WeconsidereachTwitteruserasasensor,andsetaproblemtodetectaneventbasedonsensoryobservations.LocationestimationmethodssuchasKalman“lteringandparticle“lteringareusedtoestimatethelocationsofevents.Asanapplication,wedevelopedanearthquakereportingsystem,whichisanovelapproachtonotifypeoplepromptlyofanearthquakeevent.Microblogginghasreal-timecharacteristicsthatdistin-guishitfromothersocialmediasuchasblogsandcollabo-rativebookmarks.Inthispaper,wepresentedanexampleusingthereal-timenatureofTwitter.Itishopedthatthispaperprovidessomeinsightintothefutureintegrationofsemanticanalysiswithmicrobloggingdata.9.REFERENCES[1]S.Arulampalam,S.Maskell,N.Gordon,andT.Clapp.Atutorialonparticle“ltersforon-linenon-linear/non-gaussianbayesiantracking.TransactionsonSignalProcessing,2001.[2]L.Backstrom,J.Kleinberg,R.Kumar,andJ.Novak.Spatialvariationinsearchenginequeries.InProc.,2008.[3]T.BleierandF.Freund.Earthquakewarningsystem.Spectrum,IEEE,2005.[4]K.Borau,C.Ullrich,J.Feng,andR.Shen.Microbloggingforlanguagelearning:Usingtwittertotraincommunicativeandculturalcompetence.InProc.ICWL2009,pages78…87,2009.[5]d.boyd,S.Golder,andG.Lotan.Tweet,tweet,retweet:Conversationalaspectsofretweetingontwitter.InProc.HICSS-43,2010.[6]M.EbnerandM.Schiefner.Inmicroblogging.morethanfun?InProc.IADISMobileLearningConference,2008.[7]D.Fox,J.Hightower,L.Liao,D.Schulz,andG.Borriello.Bayesian“ltersforlocationestimation.IEEEPervasiveComputing,2003.[8]G.GrosseckandC.Holotescu.Analysisindicatorsforcommunitiesonmicrobloggingplatforms.InProc.eLSEConference,2009.[9]J.HightowerandG.Borriello.Locationsystemsforubiquitouscomputing.IEEEComputer,34(8):57…66,August2001.[10]J.HightowerandG.Borriello.Particle“ltersforlocationestimationinubiquitouscomputing:Acasestudy.InProc.UbiComp04,2004.[11]B.HubermanandD.R.F.Wu.Socialnetworksthatmatter:Twitterunderthemicroscope.FirstMonday14,2009.[12]B.Jansen,M.Zhang,K.Sobel,andA.Chowdury.Twitterpower:tweetsaselectronicwordofmouth.JournaloftheAmericanSocietyforInformationScienceandTechnology,2009.[13]A.Java,X.Song,T.Finin,andB.Tseng.Whywetwitter:Understandingmicrobloggingusageandcommunities.InProc.Joint9thWEBKDDand1stSNA-KDDWorkshop2007,2007.[14]T.Joachims.Textcategorizationwithsupportvectormachines.InProc.ECMLÕ98,pages137…142,1998.[15]J.Leskovec,L.Adamic,andB.Huberman.Thedynamicsofviralmarketing.InProc.ACMConferenceonElectronicCommerce,2006.[16]Y.MatsuoandH.Yamamoto.Communitygravity:measuringbidirectionaleectsbytrustandratingononlinesocialnetworks.InProc.WWW2009,2009.[17]Q.Mei,C.Liu,H.Su,andC.Zhai.Aprobabilisticapproachtospatiotemporalthemepatternminingonweblogs.InProc.WWWÕ06,2006.[18]S.Milstein,A.Chowdhury,G.Hochmuth,B.Lorica,andR.Magoulas.Twitterandthemicro-messagingrevolution:Communication,connections,andimmediacy.140charactersatatime.OReillyMedia,[19]M.Naaman,J.Boase,andC.Lai.Isitreallyaboutme?Messagecontentinsocialawarenessstreams.InProc.CSCWÕ09,2009.[20]A.Passant,T.Hastrup,U.Bojars,andJ.Breslin.Microblogging:Asemanticanddistributedapproach.Proc.SFSW2008,2008.[21]Y.RaimondandS.Abdallah.Theeventontology,2007.http://motools.sf.net/event/event.html.[22]T.Rattenbury,N.Good,andM.Naaman.Towardsautomaticextractionofeventandplacesemanticsfrom”ickrtags.InProc.SIGIR2007,2007.[23]E.Scordilis,C.Papazachos,G.Karakaisis,andV.Karakostas.Acceleratingsesmiccrustaldeformationbeforestrongmainshocksinadriaticanditsimportanceforearthquakeprediction.JournalofSeismology,8,2004.[24]P.Serdyukov,V.Murdock,andR.vanZwol.Placing”ickrphotosonamap.InProc.SIGIR2009,2009.[25]M.Weiser.Thecomputerforthetwenty-“rstcentury.ScientiÞcAmerican,268(3):94…104,1991.