frequencybnb998.com

1CampaignOptimizationthroughBehavioralModelingandMobileNetworkAnalysisYanivAltshuler1,,ErezShmueli2,,GuyZyskind3,OrenLederman3,NuriaOliver4,Alex(Sandy)Pentland3!
Abstract—Optimizingtheuseofavailableresourcesisoneofthekeychallengesinactivitiesthatconsistofinteractionswithalargenumberof"targetindividuals",withtheultimategoalof"winning"asmanyofthemaspossible,suchasinmarketing,serviceprovision,politicalcampaigns,orhomelandsecurity.
Typically,thecostofinteractionsismonotonicallyincreasingsuchthatamethodformaximizingtheperformanceofthesecampaignsisrequired.
Inthispaperweproposeamathematicalmodeltocomputeanoptimizedcampaignbyautomaticallydeterminingthenumberofinteractingunitsandtheirtype,andhowtheyshouldbeallocatedtodifferentgeographicalregionsinordertomaximizethecampaign'sperformance.
Wevalidateourproposedmodelusingrealworldmobilitydata.
Keywords—MobileNetworks,NetworkOptimization,Marketing,Behav-iorModeling1INTRODUCTIONINaworldoflimitedresources,behaviorchangecampaigns(e.
g.
marketing,serviceprovision,politicalorhomelandsecurity)canrelyoncreativityand"coolness"uptoacertainpoint.
Thesuccessofacampaigncangenerallybedenedastheproductofreach(portionofthepopulationexposedtothecampaignmessages)andvalueofasingleinteraction(theca-pacityofamessagetoinduceacertainbehaviorinanexposedaudience).
Hence,campaignmanagerstypicallydistributetheirbudgetbetweencontentenhancement(toincreasethevalueasingleinteraction)andwidereach.
Yet,todateitseemsthattheoptimtrade-offbetweenthesetwofactorsisfoundasaresultof"intuition"ratherthanbasedonwellestablishedanalysis.
Inthispaper,weproposeanovelmathematicalmethodthat,giventhecharacteristicsofthetargetaudienceanditsabilitytobepersuaded,generatesanoptimizedcampaignstrategyintermsof:(a)thequantityofinteractingunits,alsoreferredtoasinsertionsand(b)themonetaryallocationtoeachunit.
Themodeltakesintoaccountthepopulation'smobilityinanurbanenvironmentasitcanbeinferredfromrealdatareceivedfromalargemobilephonecarrier.
Eventhough1AthenaWisdom,yaniv@athenawisdom.
com2Tel-AvivUniversity,DepartmentofIndustrialEngineering,shmueli@tau.
ac.
il3MITMediaLab,{guyz,orenled,sandy}@media.
mit.
edu4TelefonicaResearch,nuriao@tid.
esIndicatesequalcontributiondifferentpopulationslocatedindifferentenvironmentswouldbetailoredwithdifferentcampaignstrategies,theoptimalityofeachstrategywouldbemaintained.
Amajorcontributioninouroptimizationmodelistheuseofnetworkanalysismethodstoapproximatethereachofacampaign.
Morespecically,giventhenetworkofmobilitybetweenthedifferentgeographiclocations,andasubsetoflocations,weusetheGroupBetweennessCentrality(GBC)[54]–anetworkmeasurethatcalculatesthepercentageofshortestpathsamongallpairsofnetworknodesthatpassthroughapre-denedsub-setofthenetwork'snodes–toapproximatethereachofthissubsetoflocations.
WethendemonstratethatthisfunctioncanbeapproximatedusingasmoothandeasilyanalyzedGompertzfunction.
Thistacklesthemainlimitationofworksoncampaignoptimizationhitherto–efcientlyestimatingthecampaignreachasafunctionofthenumberofunitsandtheirlocations.
Finally,wevalidateourcampaignoptimizationmodelusingareal-worldmobilitynetworkinferredfromCDRdata,anddemonstratehowGBCbaseddeploymentofcampaignunitsoutperformsseveralcommonalternatives.
Therestofthispaperisorganizedasfollows.
RelatedworkisdiscussedinSection2.
AcharacterizationofacampaignmodelanditstargetoptimizationfunctionarepresentedinSection3.
Ananalyticaloptimizationofthecampaign'smodelisshowninSection4.
AvalidationofthemodelusingrealworldmobiledataisdescribedinSection5.
ConcludingremarksappearinSection6.
2RELATEDWORKInrecentyearsthesocialscienceshavebeenundergoingadigitalrevolution,heraldedbytheemergingeldof"compu-tationalsocialscience".
Lazer,Pentlandet.
al[75]describethepotentialofcomputationalsocialsciencetoincreaseourknowledgeofindividuals,groups,andsocieties,withanun-precedentedbreadth,depth,andscale.
Computationalsocialsciencecombinestheleadingtechniquesfromnetworkscience[19],[88],[116]withnewmachinelearningandpatternrecognitiontoolsspecializedfortheunderstandingofpeople'sbehaviorandsocialinteractions[50].
Marketingcampaignsareessentialfacilityinmanyareasofourlives,andspecicallyinthevirtualmedium.
Oneof2themainthruststhatpropelstheconstantexpansionsandenhancementofsocialnetworkbasedservicesisitsimmenseimpactonthe"realworld"inavarietyofeldssuchaspolitics,traditionalindustry,currencyandstocktradingandmore.
Thiseldisbecomingincreasinglypopular[51],[78],duetothepossibilityofincreasingtheimpactofcampaignsbyusingnetworkrelatedinformationinordertooptimizetheallocationofresourcesinthecampaign.
Thisreliesontheunderstandingthatasubstantialimpactofacampaignisachievedthroughthesocialinuenceofpeopleononeanother,ratherthanpurelythroughtheinteractionofcampaignmanagerswiththepeoplethatareexposedtothecampaignmessagesdirectly.
Aconstantlygrowingportionofcommercialandgovernmentmarketingbudgetsisbeingallocatedtoadvertisinginsocialplatformsthemaingoalofwhichistosparkviralphenomenathatbyspreadingthroughthesocialnetworkswouldresultinglobal"trends".
2.
1Coverageoptimization–theoryandmethodsThestudyofoptimalcoverageinpre-denedregions,byadistributedsystemofinteractingunits,hasbeenthetopicofmanyworksinthepastcoupleofdecades.
Theworkof[25]considerstheproblemoflocatingtheminimumnumberofsensorsonthenetworknodesinordertodeterminearcowvolumesoftheentirenetwork(avariantthatconsidersdynamicenvironmentsisdiscussedin[97]).
Inuencingthebehavioroflargeconsumerspopulationthroughpricesmanipulationcampaignsbygovernmentagenciesisdiscussedin[38].
In[79]amodelforoptimizingcoverageusingamultitudeofunits(intheformofsensors,fortrafcsurveillancepurposes)wasdiscussed.
Ingeneral,mostoftheanalytictechniquesusedforguaran-teeingmaximalinteractionusingdistributedactors,or"cam-paignunits"usesomesortofcellulardecompositionofthere-giontobecoveredthroughthecampaign.
Forexample,in[30]adecompositionmethodisbeingusedwhichisanalyticallyshowntoguaranteeacompletecoverageofanarea.
Anotherinterestingworkispresentedin[3],discussingtwomethodsforefcientcoveragecampaignusingmobileunits(e.
g.
carswithposters,ormobileadvertisingZeppelins),oneprobabilisticandtheotherbasedonanexactcellulardecomposition.
Similarresultscanbefoundin[12],[13].
2.
1.
1DiffusionoptimizationAnalyzingthespreadingofinformationhasbeenthefocusofmanysocialnetworksstudiesforthelastdecade[67][77].
ResearchershaveexploredboththeofinenetworksstructurebyaskingandincentivizinguserstoforwardrealmailsandE-mails[45],andonlinenetworksbycollectingandanalyzingdatafromvarioussourcessuchasTwitterfeeds[72].
Researchersbelievethatsuchtechniquescanhelpunder-standtheinter-inuenceofindividualsinnowadaysentan-gledworld,comprisingmulti-layersofsocialandmedianet-works[35],andthatitcaneventuallyleadtoaccuratepredic-tionandactiveoptimizationandconstructionforsuccessfulandlow-costviralmarketcampaigns,suchastheDARPAChallenge[94].
However,theinformationdiffusionprocessonsocialnetworksisoverwhelminglycomplicated:theoutcomeisclearlysensitivetomanyparametersandmodelsettingsthatarenotentirelywellunderstoodandmodeledcorrectly.
Asaresult,accuratetrendpredictionandinuencediffusionoptimizationarecurrentlyamongthecentralresearchtopicsintheeld.
Thedramaticeffectofthenetworktopologyonthedynam-icsofinformationdiffusionincommunitieswasdemonstratedinworkssuchas[36][90].
Oneofthemainchallengesasso-ciatedwithmodelingofbehavioraldynamicsinsocialcom-munitiesstemsfromthefactthatitofteninvolvesstochasticgenerativeprocesses.
Whilesimulationsonrealizationsfromthesemodelscanhelpexplorethepropertiesofnetworks[63],atheoreticalanalysisismuchmoreappealingandrobust.
Theidentityandcompositionofaninitial"seedgroup"intrendsanalysishasalsobeenthetopicofmuchresearch.
Kempeetal.
appliedtheoreticalanalysisontheseedsselectionproblem[70]basedontwosimpleadoptionmodels:LinearThresholdModelandIndependentCascadeModel.
Recently,Zamanetal.
developedamethodtotracerumorsbackinthetopologicalspreadingpathtoidentifysourcesinasocialnetwork[105],andsuggestthatmethodscanbeusedtolocateinuencersinanetwork.
Somescholarsexpresstheirdoubtsandconcernsfortheinuencer-drivenviralmarketingapproach,suggestingthat"everyoneisaninuencer"[18],andcompanies"shouldnotrelyonit"[115].
Theyarguethatthecontentofthemessageisalsoimportantindeterminingitsspreads,andlikelytheadoptionmodelwewereusingisnotagoodrepresentationforthereality.
2.
1.
2AdoptionmodelandsocialdiffusionAfundamentalbuildingblockintrendspredictionthatisnotyetentirelycleartoscholarsistheadoptionmodel,modelingindividuals'behaviorbasedonthesocialsignalstheyareexposedto.
Centolahasshownboththeoreticalandempiricallythatacomplexcontagionmodelisindeedmoreprecisefordiffusion[32],[33].
Usingsocialinuencerelationsderivedfromonlinedomainswasdiscussedforexamplein[91].
Muchresearchconcerningthepredictionofusers'behaviorbasedonthedynamicsintheircommunityhasbeencarriedoutinthepast,usingavarietyofapproachessuchassociologicalmethods[58],[65],communities-orientedapproaches[66],gametheory[34]andvariousmachinelearningmethods[92].
Differentadoptionmodelscandramaticallyalterthemodeloutcome[46].
Infact,arecentworkonstudyingmobileapplicationdiffusionsusingmobilephonesdemonstratedthatinrealworldthediffusionprocessisafarmorecomplicatedphenomenon,andamorerealisticmodelwasproposedin[93].
2.
2UsingmobilephonesdataforsocialsystemsmodelingTheuseofmobilephonedataforthemobilityandbehavioralmodelingoflargepopulationhasbecomepopularintherecentdecade.
In[74]thebehaviorandsocialpatternsof2.
5millionmobilephoneusers,making810millionphonecalls,wereanalyzedandresultedinefcientmappingofusers'mobilityandhousingpatterns.
Similarresultappearsin[61],3Inanotherexample,itwasshownthatthepenetrationofcellularphonestotheIsraelimarketisveryhigh,eventolowerincomehouseholds,andspeciallyamongindividualsintheagesof10to70(themainfocusoftravelbehaviorstudies)[23].
Thiswidespreaduseofcellularphonesenablesthecollectionofaccuratemobilitydatathatcanbeusedtoanalyzeandoptimizecoverageandmonitoringcampaigns.
Forexample,thisdatawasshownin[23]and[118]toprovideahighqualitycoverageofthenetwork,tracking94%ofthetrips(denedasatleast2kminurbanareas,andatleast10kminruralareas).
Theresultingdatacontainedawealthoftrafcpropertiesforanetworkofover6,000nodes,and15,000directedlinks.
Inaddition,thenetworkwasaccompaniedwithanOriginDestination(OD)matrix,specifyingstartandendpointsoftrips.
2.
3CampaignoptimizationstudiesWheneveracompanywishestointroduceanewproduct,increaseitsmarketshareormerelyretainthecurrentone,itneedstoengageitselfinmarketingefforts.
Infact,theglobalmarketingspendhasbeenrisingfastforseveraldecades,andiscurrentlyestimatedat1trilliondollarsayear,whichmakesitbetween1to2percentofglobalGDP[56].
Oneimportantdecisionwithregardtomarketinginvolvesndingtheoptimalbalancebetweencostandeffectivenessofthemarketingcampaign.
Anappropriateoptimizationmethodcanhelpeithertoobtainmoreeffectivemarketingresultsforagivenbudgetortoreducethemarketingcost.
Theadvertisingbudgetingproblemhasbeenaddressedinliteraturefromdifferentperspectives.
Earlymodelswererela-tivelysimple.
[87]focusedontherelationshipbetweencurrentmarketingspendingandfuturedemand.
In[112],VidaleandWolferepresentedsalesresponseusingthreeparameters-salesdecay,saturationlevelandaresponseconstant.
BuildingupontheVidale-Wolfemodel,[104]usedoptimalcontroltheorytoobtainanoptimaladvertisingstrategyand[43]extendedittoincludecompetitioninaduopoly.
[81]reviewedaggregateadvertisingmodels-functionsthatshowtherelationshipbetweenproductsalesandadvertisingspendingforamarketasawhole.
Hestressedthatmostmodelsatthetimeoftencontradictedoneanotherandmissedkeycomponents,makingitdifculttoputthesemodelsintopracticaluse.
Severalworks,suchas[98]and[52],suggestedthatatleasttheshorttermresponsetoadvertisingisS-shaped.
Meaningthatanincreaseinadvertisingistypicallyfollowedbyaperiodofdiminishingreturns.
In[84],MesakandHaniprovideoligopolisticjusticationforapulsingadvertisingpolicywhereS-shapedresponsefunctionsarepresent.
[43]determinestheoptimaladvertisingexpendituresforaduopolyinanequilibrium.
[99]characterizedthebrands'choiceandNashequilibriumadvertisingexpendituresinanoligopoly.
[59]showsamodelofoligopolisticcompetitioninwhichadvertisingentersintothedemandfunctionsofrms,resultinginapositiverelationshipbetweenproductpriceandthedegreeofadvertisingcooperativeness.
[89]incorporatedrandomdemandforaproducttoshowhowdemanduncertaintyaffectsadvertisingdecisions.
[68]showedthatwhenthemarketissaturated,abrandshouldchooseadefensivestrategy,inwhichthegoalispreservingtheexistingcustomersandsales.
[22]explainshowtochoosebetweengenericadvertisingandbrandadvertisingstrategiesinadynamicduopoly.
Severalresearcherssoughtmethodsofhelpingmarketingmanagersallocateagivenbudgetandoptimizeresponse.
[83]focusedonassistingmarketingmanagersinoptimizingadvertisingbudget.
Theypresentedtherstcomprehensiveallocationmodelthat,givenaxedbudget,ndstheoptimalspreadovertimeandmarketsegments.
Usingsimpleinputs,themodelcreatedby[101]efcientlyselectadvertisingsched-ulesfornetworktelevision.
Insteadoflookingatadvertisingasanexpense,[40]arguesthatadvertisingshouldberegardedasaninvestmentwithlong-livedeffects.
Theythenlayoutanapproachforcalculatingthelevelofspendingthatgeneratesoptimalreturn.
Followingthisapproach,[41]suppliestheformulasrequiredforcalculatingthelevelofspendingthatmaximizesROI.
[85]usesaMarkovdecisionprocesstoformulateastochastic,sequentialmodelthattakesintoaccountthematurityandpastadvertisingoftheproductanddeterminestheoptimaladvertisingspending.
[73]considerstheadvertisinginvestmentinaspatialmonopoly,contrastingthesociallyoptimalbehaviorofabenevolentplanneragainstthatofaprotseekingmonopolist.
Theformulasin[41]provideasolutionforasingleproduct,singlemediumandasinglestage.
[48]addressesthemulti-productadvertisingbudgetingproblem,inwhichthecross-effectsoftheadvertisedproductsaretakenintoconsideration,aswelltheeffectofthepromotedproductsontherestoftheproductsportfolio.
[106]andothersstudiedmultistageadvertisingbudgeting,showingtheshort-termandlong-termimpactondemand.
Additionalworkhasbeendoneby[24],whoaddressedthemultistagemultiproductadvertisingbudgetingproblem-optimizingthebudgetandallocationformultipleproducts,multiplesaleattributesovermultipleperiods.
Theytestedtheirmodelinanactualcampaignandobservedaclearincreaseinprotscomparedtootherapproaches.
Inthesecondpartoftheirwork,[24],theyintroduceastochasticoptimizationmodelandcompareitwiththedeterministicmodelpresentedontheirearlywork.
Sofar,wehavereviewedmodelsusedforndingoptimalexpendituresandcoverageinmarketingcampaign.
However,avariationoftheseproblemscanbefoundinotherdomainsaswell.
Forexample,in[2],theauthorsdiscusstwomethodsforefcientcoverageusingmobileunits,oneprobabilisticandtheotherbasedonanexactcellulardecomposition.
Similarresultsandmethodscanbefoundin[12],[14].
Theworkof[26]considerstheproblemoflocatingtheminimumnumberofsensorsonthenetworknodesinordertodeterminearcowvolumesoftheentirenetwork.
Thedynamicenvironmentsvariantisdiscussedin[97].
Thestudyofthecorrelationbe-tweentopologicalfeaturesofamarketingenvironmentnetworkandtheefciencyofadistributedgroupof(mobile)unitsisdiscussedin[16].
In[80]amodelforoptimizingcoverageusingamultitudeofunits(intheformofsensors,fortrafcsurveillancepurposes)wasdiscussed.
Ingeneral,mostofthe4analytictechniquesusedforguaranteeingmaximalinteractionusingdistributedactors,orcampaignunitsusesomesortofcellulardecompositionoftheregiontobecoveredthroughthecampaign.
Forexample,in[31]adecompositionmethodisbeingusedwhichisanalyticallyshowntoguaranteeacompletecoverageofanarea.
Thesemethodsaretheequivalentoftheoptimizationmethodsusedforndingtheoptimalcoverageusingvariousmedia(e.
g.
carswithposters,ormobileadver-tising,Zeppelins)intheworldofmarketingcampaigns.
2.
4CampaignoptimizationinpracticeTheauthorof[82]pointedoutthatamodelthatistobeusedbyamanagershouldbesimple,robust,easytocontrol,adaptive,ascompleteaspossibleandeasytocommunicatewith.
Inaccordancewiththisrecommendation,themodelsthatwesurveyinthissectionaresimplebutrealisticenoughtobeusedintheadvertisingindustry.
2.
4.
1EstimatingeffectivenessInthepast,short-termmeasurementofadvertisingeffective-nesswasextremelyproblematicbecauseoftheinherentlylong-termnatureofadvertisingimpactandtheverysmallshort-termeffectsofadvertising[1],[29],[110],[111].
Thus,asapracticalmatter,themediaplanneroftenemployedaproxyforadvertisingeffectiveness[41].
Popularchoicesforsuchaproxyincludereach[41],effectivereach[4],[39],[86]andaveragefrequency[41].
Reachistheproportionofthetargetaudienceexposedtoatleastoneinsertionoftheadvertisement.
Effectivereachistheproportionofthetargetaudienceexposedtoatleastthreeinsertionsoftheadvertisement.
Frequencyistheaveragenumberoftimesapersonfromthereachaudienceisexposedtoanadvertisement.
ExposuretoanadvertisementisoftenmeasuredinGrossRatingPoints(GRPs):theproductofreachandfrequency.
Forexample,100GRPscouldmeanthat100%ofthemarketisexposedoncetoanadvertisementorthat50%ofthemarketisexposedtwice[62].
Accordingto[21],itisusuallypreferredtomeasureadvertisinginGRPsandnotindollarssince:(1)mostmanagersevaluatetheeffectivenessoftheircampaignsintermsofdemandgeneratedperGRP;and(2)itisnotclearhowmuchadvertisingexposurecanbepurchasedforagivenbudget,andthusGRPsprovideaclearerpictureofadvertisinginput.
Wedenotethereachforkinsertionsasrk,andforgGRPsasrg.
Similarly,wedenotetheeffectivereachforkinsertionsaserk,andforgGRPsaserg.
Traditionally,advertisingcampaignsarequantitativelyde-scribedbytheexposuredistribution(ED),denedastheprobabilityofexposuretonone,one,uptoalloftheadsinthecampaign[42],[102].
DenotingtheexposurerandomvariableasX,rangingfrom0,1.
.
.
k,wherekisthetotalnumberofinsertionsinthecampaign.
Itwasfoundin[76]thatthemostfrequentlyusednonproprietarymodelforXisthebeta-binomialdistribution(BBD),withmassfunction:fX(x)=kx·Γ(α+β)·Γ(α+x)Γ(α+β+k)·Γ(α)·Γ(β+kx)Γ(β)whereΓisthegammafunctionandα,β>0.
Theauthorsof[41]showedthatundertheBBDmodel,thereachfunctioncanbemodeledas:rk=1k1j=0β+jα+β+jandthenumberofGRPscanbemodeledas:g=100k·αα+βThatis,thenumberofGRPsgdependslinearlyonthenumberofinsertionsk.
Atrst,thisresultmayseemstrange,asthesamevalueofGRPsmayhaveseveralcorrespondinginsertionsvalues.
Inreality,however,foramoderatelylargenumberofinsertions,theGRPs-insertionscurveisquiteat.
Thismeansthat,althoughtherearemanypossibleinsertionsvaluesforthesameGRPs,theuctuationinGRPsforthesecombinationsaresmall[109].
Inrecentyears,ithasbecomesignicantlyeasiertoestimatetheeffectivenessfunctiondirectly.
WearelivingtheeraofBigData,wherecompaniesgatherandmanagehugedatabases.
Byusingmarketresponsemodelswecantransformthisrawmarketinginformationinto'readytouse'information[62].
Asaconcreteexample,[24]modelsthesalesduetoadvertisingasafunctionofthenumberofGRPs,denotedassalesg.
Regardlessofwhethertheeffectivenessfunctionfgisesti-mateddirectlyorviaaproxy,inthesingleproductscenario,itcorrespondstoanincreasingconcavefunctionwhichmodelsdiminishingreturns[62].
Theauthorsof[41]arguethattheeffectivenessproxiesrgandergcanbewellapproximatedbyafunctionofthefollowingform:fg≈1γgλwhereγ,λ>0andg>g0forsomelowerthresholdvalueofGRPs.
Theauthorsof[62]arguethatabetterchoiceforapproxi-matingfgisthesocalled'modiedexponential'function:fg≈γ(1eδg)Inthemultiproductscenario,however,thereexistcrosselasticitiesamongtheproductsduetorelationshipsofcomple-mentarityorsubstitution[48].
Inthecaseofcomplementarity(positiveelasticity),advertisingononeproductincreasessalesofanotherproductandthiscrosseffectcanbemodeledbyanincreasingconcavefunction.
Inthecaseofsubstitution(neg-ativeelasticity),however,advertisingononeproductreducessalesoftheotherproduct(thiscrosseffectisknownascanni-balization[24]).
Thecannibalizationeffectcanbemodeledbyadecreasingconvexfunction.
Ifthisfunctionisstrictlyconvex,theresultingeffectivenessfunctionmaynotbeconcave.
Asiswellknown,concavityoftheobjectivefunctionisadesirablepropertyinamaximizationproblemsinceitguaranteesglobaloptimality(assumingaconvexfeasibledomain).
Havingsaidthat,thecrossproducteffectsareusuallysmallrelativetothe5directadvertisingeffects,andtherefore,inmostcases,itisreasonabletoassumethattheeffectivenessfunctioncanbemodeledbyanincreasingconcavefunction.
InSection4,wesuggestamethodtoderivethereachfunctionusingreal-worlddata,andproposetomodelitus-ingaGompertzfunctioninsteadofa'modiedexponential'function.
2.
4.
2Estimatingcosttheauthorsof[40]haveproposedamathematicalrelationshipbetweencostandGRPs.
Thefunctionemployedmakestwoassumptions:(1)moreGRPscostmorethanfewerGRPsandthuscostisamonotonicallyincreasingfunctionofGRPs;and(2)buyingalargenumberofGRPscanresultindiscountsandthusthecostcurveisconcave.
Theauthorsfurthersuggestthefollowingexiblefunctionalformtomodelthecostfunctioncg.
cg=C·gδwhereC>0isaconstantand0≤δ≤1isaparameterwhichreectstheexpecteddiscountingextent.
Intheremainderofthispaperweassumethatnodiscountingoccursandthusδ=1.
cg=C·g2.
4.
3OptimalcriteriaTheauthorsof[40]and[41]categorizespendingcriteriaintopopularadhoccriteria,andoptimalcriteria,whicharebasedonmodelingandsolvinganoptimizationproblem.
Althoughcriteriathatareactivelyusedinpracticeoftentendtobeintheadhoccategory,wefocushereontheoptimalcategory.
Thethreeprimaryapproachesforoptimizingthelevelofmarketingspendingare:(1)maximizingadvertisingprotabil-ity,(2)maximizingadvertisingproductivity(efciency),and(3)maximizingthereturnoninvestmentofadvertising.
MaximizingadvertisingprotabilityThisapproach,developedbyKaplanandShocker[69]startswiththeassumptionthatanadvertisingeffectivenessmeasureexists(perhapseffectivereach),andthatthismeasureisdirectlyrelatedtorevenue.
Thisseeminglystrongassumptionisnotsounreasonable,becausealmostallmediaplanners(in-cludingthemostsophisticatedones)currentlyusesurrogatesofadvertisingeffectiveness.
IfwedenoteprotabilityasE1,themeasureofadvertisingeffectivenessasfg,wheregisthenumberofGRPs,andcgisthecostofbuyinggGRPs,thentheprotabilityis:E1=K·fgcgwhereKisthedollarvalueofoneunitofeffectiveness.
Maximizingadvertisingproductivity(efciency)Maximizingproductivityorefciencyisanotheralternative.
Economistsoftenemphasizetheimportancetotheeconomyofincreasingproductivity,andmanagementscholarsrecognizethatthegreatestpotentialforgainsinproductivityareintheknowledgeandservicesectorsoftheeconomy[49].
Usingtheabovenotations,productivity(efciency)iscal-culatedas:E2=K·fgcgMaximizingthereturnoninvestmentThisapproachwasproposedin[44].
Usingtheabovenotations,returnoninvestment(ROI)iscalculatedas:E3=K·fgcgcg=K·fgcg13CHARACTERIZATIONOFACAMPAIGNLetusdeneacampaignasanactivityconnedintimeandspacewithalimitedbudgetwhoseobjectiveistosendamessageorengagewiththemaximumnumberofindividualswhoarelocatedinthatspace.
Morespecically,basedontheconventionslaidoutintheprevioussection,letusdeneaninsertionassomekindofinducedinterventionthatisdesignedtoincentivizecertainkindsofinclinedbehaviorintheaudience.
Suchinsertionscanbeforexamplelargebillboards,humanagentswhohandcouponstopassingpedestrians,etc.
Letusalsodeneadeploymentschemeasaspecicallocationalgorithmthatforagivennumberofinsertionskoutputsasetofklocationsforsaidinsertions.
Finally,letusdenethereachfunction,rk,asthenumberofindividualsexposedtoatleastoneinsertion(giventhatexactlykinsertionsweredeployedaccordingtothegivendeploymentscheme).
Selectingprotabilityasouroptimalcriteria(seeSection2.
4.
3),thereachfunctionrgastheeffectivenessmeasurefg,andC·gasthecostfunctioncg(seeSection2.
4.
2),theOptimizedCampaignProblemaimsatmaximizingthefollowingtargetfunction:Eg=K·rgC·g(1)Giventhatk≈g(seeSection2.
4.
1),theaboveequationcanberewrittentousethenumberofinsertionskinsteadofthenumberofGRPsg.
Ek=K·rkC·k(2)Next,weareleftwithefcientlymodelingthereachfunctionrk.
Assumethatthenetworkofmobilitybetweenallpossiblelocationsisavailable(Insection5wedemonstratesuchamobilitynetworkwhichisinferredfromCDRdata).
Givenanumberofinsertionskandadeploymentscheme,thereachfunctionrkcanbewellapproximatedbycalculatingtheGroupBetweennessCentrality(GBC)[54]oftheklocations(i.
e.
nodes)whicharereturnedbythedeploymentscheme.
BetweennessCentrality(BC)standsfortheabilityofanin-dividualnodetocontrolthecommunicationowinanetwork6andisdenedasthetotalfractionofshortestpathsbetweeneachpairofverticesthatpassthroughagivennode[17],[57].
InrecentyearsBetweennesswasextensivelyappliedtoanalyzevariouscomplexnetworks[20],[108]includingsocialnet-works[103],[114],computercommunicationnetworks[55],[119],andproteininteractionnetworks[27].
Holme[64]hasshownthatBetweennessishighlycorrelatedwithcongestioninparticlehoppingsystems.
ExtensionsoftheoriginaldenitionofBCareapplicablefordirectedandweightednetworks[28],[117]aswellasformultilayernetworkswheretheunderlyinginfrastructureandtheorigin-destinationoverlayareexplicitlydened[95].
TheGBCofagivengroup(UV)ofverticesaccountsforallroutesthatpassthroughatleastonememberofthegroup.
Letσs,tbethenumberofshortest-pathroutesfromstot,andletσs,t(U)bethenumberofshortest-pathroutesfromstotpassingthroughatleastonevertexinU:GBC(U)=s,t∈V\C|s=tσs,t(U)σs,t(3)Whilethe"optimaldeploymentscheme",bydenition,wouldselectthesetofklocationsthatyieldthemaximalgroupbetweenesscentrality,ourproposedmodeldoesnotassumeanyconstraintonthedeploymentschemeandenablestheplannerstoselecttheoptimalnumberandtypeoftheunits,asafunctionofthedeploymentschemeused.
Morespecically,inmanycasestheoptimaldeploymentschememightbeunfeasible,involveadditionalcosts,orbesubjecttovariousregulatoryconstraints.
Insuchcasescampaignmanagersmaychooseadifferent,non-optimal,deploymentscheme.
Regardlessofthedeploymentschemechosen,itwouldofcoursestillprovidemonotonicallyincreasingreach(andGBC),albeitwithalowerinclineratecomparedtotheoptimalone.
Finally,wemodeltheapproximatedreachfunction,usingthewell-knownGompertzfunction[60]:rk=aebeck(4)TheGompertzfunctioniswidelyusedformodelingagreatvarietyofprocesses,(duetotheexiblewayitcanbecontrolledusingtheparametersa,bandc),suchasmobilephoneuptake[100]orpopulationinaconnedspace[53].
Itsabilitytomodeltheprogressofoptimizationprocessasafunctionoftheavailableresourcescanbeseenforexamplein[5]–[7],[15].
InSection5,wepresentempiricalevidencewhichillustrateshowthereachfunctionrkapproximatedbytheGBCofthreedifferentdeploymentschemescanbettedefcientlyintoaGompertzfunction.
AssigningthevaluesofEquation4backintoEquation2,resultsinthefollowingtargetfunction:Ek=K·aebeckC·k(5)Thecampaignwillbeoptimizedbydeterminingtheop-timalnumberofinsertions(k)andtheoptimalcost(C)ofeachindividualinsertionthatwouldmaximizethecampaign'sperformance.
4OPTIMIZEDCAMPAIGNSAtthispoint,wehaveaclearmodelforestimatingtheefciencyofacampaign,thatisdependentonthenumberofinsertionsandthecostofeachinsertion.
4.
1OptimizingtheNumberofInsertionsWenowturnourattentiontondingtheoptimalnumberofinsertionskthatwouldmaximizetheprotabilityofthecampaign.
First,weobtainthederivativetondthecriticalpointsofthefunctionweseektooptimize:Ekk=K·(aebeck)kC(6)NullifyingEquation6resultsin:(aebeck)k=CK(7)Inthatcase,usingEquation7weobtain:a·b·c·eck·ebeck=CKwhichinturnimplies:beck+cklnCa·b·c·K=0(8)Wenotethata,b,c>0.
AnalyzingEquation8wecanthenseethatincaseswhere:CK≤a·ce(9)andwhereW(x)istheLambertproductlog,thatcanbecalculatedusingtheseries:W(x)=∞n=1(1)n1nn2(n1)!
xn(10)theoptimalvalueofkwouldequal:k1=ln1a·b·c·CKW1a·c·CKck2=ln1a·b·c·CKW11a·c·CKc(11)NotethatW(x)istheLambertproductlog,andWk(x)isitsanalyticcontinuationoverthecomplexplane(thevaluesofthefunctionsW(x)andW1(x)inthesegmentimpliedbytheconstraintofEquation9areillustratedinFigure1).
Returningtotheoptimizationofthecampaign,wenowassignthevaluesoftheoptimalnumberofinsertionsofEquation11intothedenitionofEk,asfollows:Ek1=K·aebeck1k1·CEk2=K·aebeck2k2·C7Fig.
1.
TheupperandlowerchartsdepictthevaluesoftheLambertfunctionsW(x)andW1(x)inthesegment[1e,0],respectively.
Thesegmentisimpliedbythecon-straintofEquation9.
andusingthepropertiesoftheWfunction,simplifyitintothefollowingform:Ekmax=max{Ek1,Ek2}where:(12)Ek1=a·b·K·γ·W(b·γ)+1W(b·γ)ln(γ)Ek2=a·b·K·γ·W1(b·γ)+1W1(b·γ)ln(γ)wheretheCampaignBenetFactorγisdenedas:γ=1a·b·c·CKFromEquation12weseethattheoptimizationofacam-paignisafunctionoftheCampaignBenetFactorγ,whichtakesintoaccountthedollarvalueofoneunitofeffectiveness,thetotalcost,aswellasthedeploymentscheme(characterizedbythevaluesofa,bandc).
NoticethatthevalueofCdoesnotaffectthevaluesofa,bandc(astheyaresolelyderivedfromthecoverageefciencyofthemobilitypatterns).
4.
2OptimizingtheCostofaSingleInsertionWenowproceedtondingtheoptimaltypeofunitsthatshouldbedeployed,byoptimizingEquation12withrespecttothecostofvariouspossiblekindsofinsertions.
Denition1:LetCostBasedenotethecostofthe"mostexpensive"insertionswhichwewillassumearetheunitsthatprovidethehighestvaluetotheinitiatorsofthecampaign,namely—thoseinsertionsthatpersuadethemaximalamountofindividualstotakethedesiredaction.
WeherebydeneK,thedollarvalueofoneunitofeffec-tiveness,asafunctiondependentontheproportionbetweenthecostofthegiventypeofinsertionsandthecostoftheoptimal,butmostexpensive,insertions:K=fSCCostBase(13)Wenowlookintondingtheoptimalcostofinsertionsthatwouldmaximizetheprotabilityofacampaign.
Todoso,werstreviseEquation12inordertotakeintoaccountthedifferenttypesofinsertions:(14)Ek1=a·b·fS·γ·W(b·γ)+1W(b·γ)ln(γ)Ek2=a·b·fS·γ·W1(b·γ)+1W1(b·γ)ln(γ)andwhere:γ=1a·b·c·CfSWethenmaximizethenancialmeritsofthecampaign(namely,max{Ek1,Ek2}),bycalculatingthepartialderiva-tivesEk1CandEk2C:Ek1C=(15)1c·W(b·γ)ln(γ)+1fSC·CCostBase·fSW(b·γ)and:Ek2C=1c·W1(b·γ)ln(γ)+1fSC·CCostBase·fSW1(b·γ)Onceagain,bynullifyingthepartialderivativeandndingthecriticalpoints,weobtainthefollowingsetofequations:(16)0=W(b·γ)ln(γ)+1CCostBase·fSCCCostBase·fSCCostBase1W(b·γ)or:0=W1(b·γ)ln(γ)+1CCostBase·fSCCCostBase·fSCCostBase1W1(b·γ)where:γ=1a·b·c·CK·fSCCostBase18Equation16cannowbeusedtocalculatetheexactoptimalcostofasingleinsertion,foreverycost-valuerelation,andforeverydeploymentscheme!
5CAMPAIGNOPTIMIZATIONFORAREAL-WORLDMOBILITYNETWORKInthissectionwevalidateourcampaignoptimizationmodelusingareal-worldmobilitynetworkinferredfromCDRdata.
5.
1TheDatasetWeusedCallDataRecords(CDR)fromalargemobilecarriertocreateanetworkthatcapturesthetravelingpatternsamongdifferenturbanareas.
CDRarereadilyavailabletoday,astheyarecollectedbyallcarriersandinmost(ifnotall)countries.
Furthermore,theserecordsareconstantlycollectedinanautomatedmanner,thusincreasingthelikelihoodofthedatabeingobjectiveanduniformacrosslocationsandoperators.
Morespecically,wedenoteG=V,Etobetheundirectednetworkgraph,whereVisthesetofverticesrepresentingthecelltowersandEisthesetofweightededgesrepresentingtripsormovementsofpeoplebetweentwocelltowers.
Theweightofeachedgerepresentsthenumberoftripsthatpeoplemadebetweenthecelltowersconnectedbythatedge.
EachCDRcontainsananonymizedidentierofthecaller/callee,thecalltime,andthecelltowerthatthephonewasconnectedtowhenthecall(orSMS)originated.
Atripisdenedasachangeoflocationbythecaller/callee,detectedbytheexistenceoftwoconsecutivecallsfromtwodifferenttowersorbyachangeinthecelltowerduringanexistingcall.
Asshowningure2(left),theweightsoftheedgesinthenetworkseemtofollowapower-lawdistribution.
Weusedthemethodsuggestedin[37]and[113]todeterminethebestXminandcorrespondingγvaluesthattapower-lawdistribution.
Sincethefocusofthispaperisnotrelatedtothenetwork'stopology,ouranalysisdidnotincludethestatisticaltestsperformedin[37].
Figure2(right)showstheProbabilityDensityFunction(PDF)andthebestpower-lawtwithXmin=3andγ=1.
87.
Inordertofocusonthenetwork'sstructurethatrepresentsurbanmobilitypatternsoflargepopulations,weretainedonlyedgeswithweightshigherthan10(arbitrarilychosenthreshold),producingagraphwith|V|=18,315and|E|=130,313.
TheComplementaryCumulativeDistributionFunction(CCDF)ofnodedegreesisshowninFigure3(left).
Onceagain,weusedthemethodsuggestedby[37]and[113]tondtheparametersthatbesttapower-lawdistribution.
Figure3(right)showsthePDFandthebestpower-lawtwithXmin=45andγ=5.
05.
5.
2OptimizedDeploymentSchemesAcquiringinformationregardingthemobilitypatternsoftheaudiencemembers,andthespecicnetworkthatisgeneratedthroughthosepatternscanbeusedtosubsequentlyderiveoptimizedlocationsforthecampaignunits,ortotheveryleast,provideawayofmeasuringtheutilizationofasetoflocations,bycalculatingtheGBCoftheset,andcomparingittothatoftheoptimalone.
SeveralcombinatorialoptimizationtechniquescanbeusedtondagroupofnodesofgivensizethathasthelargestGBC,includinggreedyapproximation[47],aclassicalDepthFirstBranchandBound(DFBnB)heuristicsearchalgorithm[71],ortherecentlyproposedPotentialSearch[107].
BoththeDFBnBandthePotentialalgorithmsareanytimesearchalgorithms[120],meaningthattheirexecutioncanbestoppedatanypointoftime,yieldingthebestsolutionfoundsofar.
Similarlyto[96],inthissection,weexaminethreemethodsforndingthegroupofsizekwiththehighestGBC,andapplythemonourreal-worldmobilitynetwork.
Therstmethod,RandomDeployment,simplyselectsarandomsetofkvertices.
Thesecondmethod,BCDeployment,choosesthekverticeswiththehighestindividualBetweenessCentralityscores.
Thethirdmethod,GBCDeployment,usesagreedyalgorithmwhichiterativelyndsthevertexwhichimprovesthegroup'sGBCscorethemostandaddsittothegroup.
DuetotheruntimecomplexityoftheGBCcalculationandtheGBCdeploymentscheme,weperformedtwostagesofsamplingourmobilitynetwork.
First,asubsetof5000verticeswasrandomlyselected.
Then,weretainedonlythelargestconnectedcomponent,denotedbyG,anetworkcontaining991verticesand3,304edges.
Anillustrationoftheresultingnetworkisshowningure4.
Fig.
4.
NetworkG'Itcanbeseenthatthisnetworkisagoodsamplingoftheoriginalnetwork,asthemediandistancefromeachtrimmednodetotheclosestnodeinthesamplednetworkis4,whilethediameterofthegraphitselfis22.
Figure5(left)showsthedistributionofdistancesoftrimmedverticesfromnodesinthesamplednetworkandFigure5(right)depictsthePDFandthebestpower-lawtwithXmin=3andγ=2.
08.
Weexaminedthethreedeplymentschemesvariants,calcu-9Fig.
2.
EdgeweightsinnetworkG.
Theleftgureshowsdistributionofedgeweights.
TherightguredepictstheProbabilityDensityFunctionFig.
3.
DistributionofthenodedegreesofG.
TheleftguredepictstheComplementaryCumulativeDistributionFunctionforalldegrees.
TherightgureillustratestheProbabilityDensityFunctionforalldegreeshigherthanXminlatedtheircorrespondingGBCvalues,andttedthemontoaGompertzfunctionusingregression.
Figure6illustratesthethreedeploymentschemesforourmobilitynetwork.
TheirttingyieldedthefollowingGompertzregressions:RandomDeployment:rk=0.
76e3.
14e0.
02k(17)BCDeployment:rk=0.
92e0.
94e0.
2kGBCDeployment:rk=0.
96e1.
25e0.
38kTheregressionshadthefollowingtquality(intermsofR2):1)RandomDeployment-0.
66832)BCDeployment-0.
80353)GBCDeployment-0.
8531Consistentlywith[95],usingtheGBCDeploymentwesawthatitispossibletocoverthevastmajorityofthemostpopularmobilitynodeswithafewdozeninsertions.
TheBCDeploymentproducedhighqualitydeploymentsaswell,althoughbasedonourndingsitrequiredahighernumberofinsertions.
Itseemsthatbothschemessignicantlyoutperform10Fig.
5.
DistributionofdistancesbetweentrimmedverticesandnodesinthesamplednetworkG(left)andProbabilityDensityFunctionofthedistancesbetweentrimmedverticesandnodesinthesamplednetworkG(right)RandomDeployment,whichrequired100insertionstoreach50percentcoverage,andafewhundrednodestoguaranteeanear-fullone.
Itisimportanttonotethatrkcansignicantlychangefordifferentnetworks(modelingdifferenturbanenvironmentsmobilitypatterns).
Withthisinmind,wecannowproceedtondingtheoptimalnumberofinsertions,andsubsequently—themaximizationofthemonetaryutilizationoftheinvestmentinthesystem.
5.
3OptimizingthenumberofinsertionsandthecostperinsertionInthissectionwedemonstratehowtheproposedmethodcanbeusedtosignicantlyincreasetheutilizationofagivencampaign.
Forsimplicity,weassumethatbothKandthecostperunit(C)arecontinuous,andfollowsomepre-denedfunction,knowntothecampaignmanagers.
,Forexample,wemayimagineafunctionthatfollowsasub-linearcorrelation,suchasthefollowing:fS1CCostBase=CCostBase2Themeaningofthisfunctionisthatinsertionsthatcosthalfoftheoptimalinsertionspossible,wouldgenerate25%valueoftheoptimal(mostexpensive)ones.
Alternatively,wemayimagineenvironmentswherethecorrelationbetweeninsertions'costandvalueissuper-linear,convergingtoone,suchasthefunction:fS2CCostBase=CCostBaseInthisanalysisweshallusethevaluesoftheGompertzapproximationasshownaboveinEquation17:RandomDeployment:a=0.
76,b=3.
14,c=0.
02BCdeployment:a=0.
92,b=0.
94,c=0.
2GBCdeployment:a=0.
96,b=1.
25,c=0.
38Inthiscase,nullifyingthepartialderivativeofEquation15forGBCdeployment,andassumingfS1asameasurementofK,wouldyield:Ek1C=0→(18)W(1.
25·γ)=ln(γ)+1W(1.
25·γ)and:Ek2C=0→W1(1.
25·γ)=ln(γ)+1W1(1.
25·γ)subsequentlyimplying:γopt≈0.
2837(inthisexample,theoptimalvalueofγforEk2hasanon-zeroimaginarycomponent).
Usingthisoptimalvalueofγwewouldnowget:γopt=1a·b·c·CfS1=2.
19·Cost2BaseC=0.
2837andfromthiswereceive:Copt=2.
19·Cost2Base0.
2837·K≈7.
72·Cost2Base(19)11Fig.
6.
GBCforthreedeploymentschemeswithrespecttothenumberofusedunits,usingmobilitypatternsextractedfrommobilephonesdata(lefttoright):(a)RandomDeployment,(b)BCDeployment,and(c)GBCDeployment.
TheappropriateGompertztofthecurvesisalsoincluded.
FromEquation19wecanobtainforeachkindofcampaigntheoptimaltypeofinsertionsthatshouldbeused,inordertomaximizeEquation5—theoptimizationfunctionthatmaximizesacampaign'sprotability.
AssigningthisbackintoEquation11,wecangettheoptimalnumberofinsertionsforeachpotentialcampaign(seeanillustrationinFigure7):k1=(20)2.
63·W2.
74·Cost2Base+0.
7865.
26·ln(CostBase)and:k2=2.
63·W12.
74·Cost2Base+0.
7865.
26·ln(CostBase)Notethatthepreviousexampledependsontheassumptionofaquadraticrelationbetweenthecostandqualityoftheinsertions,andontheassumptionregardingthedeploymentscheme(i.
e.
theparametersoftheGompertzModel).
However,givenanysubstitutefortheseassumptions,correspondingsolutionstotheoptimizedcampaignproblemwillbegeneratedbythemodel.
RepeatingtheaboveprocesswiththeGompertzttedvaluesforBCDeployment,aswellasRandomDeployment,wecanobtaintheoptimizedcampaigns(intermsofcostperinsertion,andnumberofinsertions)forthesedeploymentschemes.
Theaboveanalysisexaminedtheprotabilityofthecam-paign(seeagainSection2.
4.
3).
Ifweareinterestedinthereturnoninvestmentofthecampaign,wecandevidetheprotabilityofthecampaignbyitscost(i.
e.
thenumberofinsertionstimesthecostperinsertion).
Figure8showsthereturnoninvestmentofthethreedeploymentschemesasafunctionoftheirprotability.
ThegureillustratesthesuperiorityoftheGBCmethod,aswellastheperformanceoftheoptimizedcampaigns,comparedtonon-optimizedones.
6CONCLUSIONSInthispaperwestudiedtheproblemofcampaignoptimization.
Westartedbyformalizingtheproblemofoptimizingacam-paignbyndingtheoptimaltrade-offbetweentheresources(cost)allocatedtoeachsingleunit,andthenumberofsuchunits.
Wehavepresentedanovelmodeltoanalyticallygenerateanoptimizedstrategyformarketingcampaigns,anddemon-stratedhowitcanbeusedwithaggregatedandanonymizedmobilitydatareceivedfrommobilecarriers.
Specically,wehaveshownawaytoanalyticallycalculatetheexactoptimalcostforunitsinacampaign,aswellastheoptimalnumberofsuchunits,thatwouldguaranteeamaximalutilizationofthecampaign'sbudget.
Inthisworkwehavediscussedtheoptimizationofcam-paignsresourcesutilization.
However,theexposureresultstobeguaranteedusingsuchresourcesandtheproposedmethodwereleftoutsidethescopeofthiswork.
Thisaspecthowever(namely,theoreticallowerboundsforanyconceivablecam-paignstrategy),wasdiscussedinworkssuchas[13]–whereagenericanalyticalboundwasdeveloped,undertheassumptionthatthecampaign'stargetwillpracticeanadversarialstrategythatwouldminimizetheirexposure,or[8],[9],[13]–wheretheimpactofchangesinthetopologicalpropertiesoftheenvironmentandthecampaign'stheoreticalutilizationwereanalyzed.
Finally,itisinterestingtonotethattheproblemofndinganoptimal(andoptionallydynamic)"engagementstrategy"isrelatedtootherkindsofmonitoringproblems,suchasmonitoringforevadinglandtargetsbyaockofUnmannedAirVehicles(UAV).
Inthisproblem,however,thefactthatthepathsoftheUAVsisunconstrained(astheyareyingintheair)makesthecalculationofanear-optimalmonitoringstrategyfairlyeasy[10],[11].
REFERENCES[1]MagidMAbrahamandLeonardMLodish.
Gettingthemostoutofadvertisingandpromotion.
HarvardBusinessReview,68(3):50,1990.
[2]ErcanAcar,YangangZhang,HowieChoset,MarkSchervish,AlbertGCosta,RenataMelamud,DavidLean,andAmyGraveline.
Pathplanningforroboticdemininganddevelopmentofatestplatform.
InInternationalConferenceonFieldandServiceRobotics,pages161–168,2001.
12Fig.
7.
AnillustrationofEquation16,undertheGBCDeploymentscheme,asapproximatedbytheGompertzfunctionmodelingofEquation17.
Forexample,whenasingleunitcosts1%ofthemaximalcampaignimpact,theoptimalnumberofcampaignunitswouldbe25,whereasifcheaperunitsareused(suchasunitsthatcostmerely12%ofthemaximalcampaignimpact)theoptimalnumberofunitswouldbe28.
Alternatively,forexpensiveunitsthatcost5%ofthemaximalcampaign'simpact,theoptimalnumberofcampaignunitswoulddropto16.
Fig.
8.
Comparativeanalysisof3deploymentschemes–GBCDeployment,BCDeploymentandRandomDeployment,intermsoftheirreturnoninvestment,asafunctionoftheirprotability.
Itcanbeseethatthereturnoninvestmentofallmethodsismonotonicallyincreasing.
However,itcanclearlybeseenthatwhereasRandomDeploymentachievesonlyaslowincreaseinreturnoninvestment,theperformanceofGBCDeploymentandBCDeploymentaremuchbetter.
[3]E.
U.
Acar,Y.
Zhang,H.
Choset,M.
Schervish,A.
G.
Costa,R.
Mela-mud,D.
C.
Lean,andA.
Gravelin.
Pathplanningforroboticdemininganddevelopmentofatestplatform.
InInternationalConferenceonFieldandServiceRobotics,pages161–168,2001.
[4]AlvinAAchenbaum.
Effectiveexposure:Anewwayofevaluatingmedia.
InAssociationofNationalAdvertisersMediaWorkshop,NewYork,1977.
[5]Y.
Altshuler,N.
Aharony,M.
Fire,Y.
Elovici,andAPentland.
Incrementallearningwithaccuracypredictionofsocialandindividualpropertiesfrommobile-phonedata.
CoRR,2011.
[6]Y.
Altshuler,N.
Aharony,A.
Pentland,Y.
Elovici,andM.
Cebrian.
Stealingreality:Whencriminalsbecomedatascientists(orviceversa).
IntelligentSystems,IEEE,26(6):22–30,nov.
-dec.
2011.
[7]Y.
Altshuler,M.
Fire,N.
Aharony,Y.
Elovici,andAPentland.
Howmanymakesacrowdonthecorrelationbetweengroups'sizeandtheaccuracyofmodeling.
InInternationalConferenceonSocialComputing,Behavioral-CulturalModelingandPrediction,pages43–52.
Springer,2012.
[8]Y.
Altshuler,I.
A.
Wagner,andA.
M.
Bruckstein.
Shapefactor'seffectonadynamiccleanersswarm.
InThirdInternationalConferenceonInformaticsinControl,AutomationandRobotics(ICINCO),theSec-ondInternationalWorkshoponMulti-AgentRoboticSystems(MARS),pages13–21,2006.
[9]Y.
Altshuler,I.
A.
Wagner,andA.
M.
Bruckstein.
Onswarmoptimalityindynamicandsymmetricenvironments.
volume7,page11,2008.
[10]Y.
Altshuler,V.
Yanovski,I.
A.
Wagner,andA.
M.
Bruckstein.
The13cooperativehunters-efcientcooperativesearchforsmarttargetsusinguavswarms.
InSecondInternationalConferenceonInformaticsinControl,AutomationandRobotics(ICINCO),theFirstInternationalWorkshoponMulti-AgentRoboticSystems(MARS),pages165–170,2005.
[11]Y.
Altshuler,V.
Yanovsky,A.
M.
Bruckstein,andI.
A.
Wagner.
Efcientcooperativesearchofsmarttargetsusinguavswarms.
ROBOTICA,26:551–557,2008.
[12]Y.
Altshuler,V.
Yanovsky,I.
Wagner,andA.
Bruckstein.
Swarmintelligencesearchers,cleanersandhunters.
SwarmIntelligentSystems,pages93–132,2006.
[13]YanivAltshulerandAlfredM.
Bruckstein.
Staticandexpandinggridcoveragewithantrobots:Complexityresults.
TheoreticalComputerScience,412(35):4661–4674,2011.
[14]YanivAltshulerandAlfredMBruckstein.
Staticandexpandinggridcoveragewithantrobots:Complexityresults.
TheoreticalComputerScience,412(35):4661–4674,2011.
[15]YanivAltshuler,MichaelFire,NadavAharony,ZeevVolkovich,YuvalElovici,andAlexSandyPentland.
Trade-offsinsocialandbehavioralmodelinginmobilenetworks.
InSocialComputing,Behavioral-CulturalModelingandPrediction,pages412–423.
Springer,2013.
[16]YanivAltshuler,IsraelAWagner,andAlfredMBruckstein.
Onswarmoptimalityindynamicandsymmetricenvironments.
economics,7:11,2008.
[17]J.
M.
Anthonisse.
Therushinadirectedgraph.
TechnicalReportBN9/71,StichtingMathematischCentrum,Amsterdam,1971.
[18]E.
Bakshy,J.
M.
Hofman,W.
A.
Mason,andD.
J.
Watts.
Everyone'saninuencer:quantifyinginuenceontwitter.
InProceedingsofthefourthACMinternationalconferenceonWebsearchanddatamining,pages65–74.
ACM,2011.
[19]Albert-LaszloBarabasiandR.
Albert.
Emergenceofscalinginrandomnetworks.
Science,286(5439):509–512,1999.
[20]M.
Barthelemy.
Betweennesscentralityinlargecomplexnetworks.
TheEuropeanPhysicalJournalB–CondensedMatter,38(2):163–168,March2004.
[21]FrankMBass,NorrisBruce,SumitMajumdar,andBPSMurthi.
Wearouteffectsofdifferentadvertisingthemes:Adynamicbayesianmodeloftheadvertising-salesrelationship.
MarketingScience,26(2):179–195,2007.
[22]FrankMBass,AnandKrishnamoorthy,AshutoshPrasad,andSureshPSethi.
Genericandbrandadvertisingstrategiesinadynamicduopoly.
MarketingScience,24(4):556–568,2005.
[23]ShlomoBekhor,YehoshuaCohen,andCharlesSolomon.
Evaluatinglong-distancetravelpatternsinisraelbytrackingcellularphonepositions.
JournalofAdvancedTransportation,pagesn/a–n/a,2011.
[24]CesarBeltran-Royo,HZhang,LABlanco,andJAlmagro.
Multistagemultiproductadvertisingbudgeting.
EuropeanJournalofOperationalResearch,2012.
[25]ConfessoreG.
GentiliM.
Bianco,L.
Combinatorialaspectsofthesensorlocationproblem.
AnnalsofOperationResearch,144(1):201–234,2006.
[26]LucioBianco,GiuseppeConfessore,andMonicaGentili.
Combina-torialaspectsofthesensorlocationproblem.
AnnalsofOperationsResearch,144(1):201–234,2006.
[27]P.
Bork,L.
J.
Jensen,C.
vonMering,A.
K.
Ramani,I.
Lee,andE.
M.
Marcotte.
Proteininteractionnetworksfromyeasttohuman.
Curr.
Opin.
Struct.
Biol.
,14(3):292–299,2004.
[28]U.
Brandes.
Onvariantsofshortest-pathbetweennesscentralityandtheirgenericcomputation.
SocialNetworks,30(2):136–145,2008.
[29]SimonBroadbent.
Pointofview-whatisasmalladvertisingelasticity.
JournalofAdvertisingResearch,29(4):37–39,1989.
[30]Z.
Butler,A.
Rizzi,andR.
Hollis.
Distributedcoverageofrectilinearenvironments.
InProceedingsoftheWorkshopontheAlgorithmicFoundationsofRobotics,2001.
[31]ZackJButler.
Distributedcoverageofrectilinearenvironments.
PhDthesis,CarnegieMellonUniversity,2000.
[32]D.
Centola.
Thespreadofbehaviorinanonlinesocialnetworkexperiment.
science,329(5996):1194,2010.
[33]D.
CentolaandM.
Macy.
Complexcontagionsandtheweaknessoflongties.
AmericanJournalofSociology,113(3):702,2007.
[34]N.
Cesa-BianchiandG.
Lugosi.
Potential-basedalgorithmsinon-linepredictionandgametheory.
MachineLearning,51:239–261,2003.
[35]M.
Cha,H.
Haddadi,F.
Benevenuto,andK.
P.
Gummadi.
Measuringuserinuenceintwitter:Themillionfollowerfallacy.
In4thInter-nationalAAAIConferenceonWeblogsandSocialMedia(ICWSM),2010.
[36]H.
Choi,S.
H.
Kim,andJ.
Lee.
Roleofnetworkstructureandnetworkeffectsindiffusionofinnovations.
IndustrialMarketingManagement,39(1):170–177,2010.
[37]AaronClauset,CosmaRohillaShalizi,andMarkEJNewman.
Power-lawdistributionsinempiricaldata.
SIAMreview,51(4):661–703,2009.
[38]M.
C.
CohenandP.
Harsha.
Designingpriceincentivesinanetworkwithsocialinteractions.
Submitted,2013.
[39]CSamuelCraigandAvijitGhosh.
Usinghousehold-levelviewingdatatomaximizeeffectivereach.
JournalofAdvertisingResearch,1993.
[40]PeterJDanaherandRolandTRust.
Determiningtheoptimallevelofmediaspending.
JournalofAdvertisingResearch,34(1):28–34,1994.
[41]PeterJDanaherandRolandTRust.
Determiningtheoptimalreturnoninvestmentforanadvertisingcampaign.
EuropeanJournalofOperationalResearch,95(3):511–521,1996.
[42]PJDanaher.
Parameterestimationandapplicationsforageneralisationofthebeta-binomialdistribution.
AustralianJournalofStatistics,30(3):263–275,1988.
[43]KennethRDeal.
Optimizingadvertisingexpendituresinadynamicduopoly.
OperationsResearch,27(4):682–692,1979.
[44]NormanKDhalla.
Assessingthelongtermvalueofadvertising.
HarvardBusinessReview,56(1):87–95,1978.
[45]P.
S.
Dodds,R.
Muhamad,andD.
J.
Watts.
Anexperimentalstudyofsearchinglobalsocialnetworks.
Science,301(5634):827,2003.
[46]P.
S.
DoddsandD.
J.
Watts.
Universalbehaviorinageneralizedmodelofcontagion.
PhysicalReviewLetters,92(21):218701,2004.
[47]S.
Dolev,Y.
Elovici,R.
Puzis,andP.
Zilberman.
Incrementaldeploy-mentofnetworkmonitorsbasedongroupbetweennesscentrality.
Inf.
Proc.
Letters,109:1172–1176,2009.
[48]PeterDoyleandJohnSaunders.
Multiproductadvertisingbudgeting.
MarketingScience,9(2):97–113,1990.
[49]PeterFDrucker.
Thenewproductivitychallenge.
QualityinHigherEducation,37,1995.
[50]N.
Eagle,A.
Pentland,andD.
Lazer.
Inferringsocialnetworkstructureusingmobilephonedata.
ProceedingsoftheNationalAcademyofSciences(PNAS),106:15274–15278,2009.
[51]NathanEagle,MichaelMacy,andRobClaxton.
Networkdiversityandeconomicdevelopment.
Science,328(5981):1029–1031,2010.
[52]JosephOEastlackandAmbarGRao.
Modelingresponsetoadver-tisingandpricingchangesforv-8cocktailvegetablejuice.
MarketingScience,5(3):245–259,1986.
[53]G.
M.
Erickson,P.
J.
Currie,B.
D.
Inouye,andA.
A.
Winn.
Tyrannosaurlifetables:Anexampleofnonaviandinosaurpopulationbiology.
Science,313(5784):213–217,2006.
[54]M.
G.
EverettandS.
P.
Borgatti.
Thecentralityofgroupsandclasses.
MathematicalSociology,23(3):181–201,1999.
[55]M.
Faloutsos,P.
Faloutsos,andC.
Faloutsos.
Onpower-lawrela-tionshipsoftheinternettopology.
SIGCOMMComput.
Comm.
Rev.
,29(4):251–262,1999.
[56]McKinseyChiefMarketing&SalesOfcerForum.
BigData,Analytics,andtheFutureofMarketing&Sales.
2013.
14[57]L.
C.
Freeman.
Asetofmeasuresofcentralitybasedonbetweenness.
Sociometry,40(1):35–41,1977.
[58]D.
Friedman,A.
Steed,andM.
Slater.
Spatialsocialbehaviorinsecondlife.
InProc.
IntelligentVirtualAgentsLNAI4722,pages252–263,2007.
[59]JamesWFriedman.
Advertisingandoligopolisticequilibrium.
TheBellJournalofEconomics,pages464–473,1983.
[60]BenjaminGompertz.
Onthenatureofthefunctionexpressiveofthelawofhumanmortality,andonanewmodeofdeterminingthevalueoflifecontingencies.
PhilosophicalTransactionsoftheRoyalSocietyofLondon,115:513–583,1825.
[61]MartaC.
Gonzalez,CesarA.
Hidalgo,andAlbert-LaszloBarabasi.
Understandingindividualhumanmobilitypatterns.
Nature,453(7196):779–782,062008.
[62]DominiqueMHanssens,LeonardJParsons,andRandallLSchultz.
Marketresponsemodels:Econometricandtimeseriesanalysis,vol-ume12.
Springer,2003.
[63]C.
P.
Herrero.
Isingmodelinscale-freenetworks:Amontecarlosimulation.
PhysicalReviewE,69(6):067109,2004.
[64]P.
Holme.
Congestionandcentralityintrafcowoncomplexnetworks.
AdvancesinComplexSystems,6(2):163–176,2003.
[65]C.
L.
HsuandH.
P.
Lu.
Whydopeopleplayon-linegamesanextendedtamwithsocialinuencesandowexperience.
InformationandManagement,41:853–868,2004.
[66]C.
L.
HsuandH.
P.
Lu.
Consumerbehaviorinonlinegamecommuni-ties:Amotivationalfactorperspective.
ComputersinHumanBehavior,23:1642–1659,2007.
[67]B.
A.
Huberman,D.
M.
Romero,andF.
Wu.
Socialnetworksthatmatter:Twitterunderthemicroscope.
FirstMonday,14(1):8,2009.
[68]JohnPhilipJones.
HowMuchisEnough:GettingtheMostfromYourAdvertisingDollar.
LexingtonBooks,1992.
[69]RobertSKaplanandAllanDShocker.
Discounteffectsonmediaplans.
JournalofAdvertisingResearch,11(3):37–43,1971.
[70]D.
Kempe,J.
Kleinberg,andE.
Tardos.
Maximizingthespreadofinuencethroughasocialnetwork.
InProceedingsoftheninthACMSIGKDDinternationalconferenceonKnowledgediscoveryanddatamining,pages137–146.
ACM,2003.
[71]R.
E.
KorfandW.
Zhang.
Performanceoflinear-spacesearchalgorithms.
ArticialIntelligence,79(2):241–292,1995.
[72]H.
Kwak,C.
Lee,H.
Park,andS.
Moon.
Whatistwitter,asocialnetworkoranewsmediaInProceedingsofthe19thinternationalconferenceonWorldwideweb,pages591–600.
ACM,2010.
[73]LucaLambertini.
Advertisinginadynamicspatialmonopoly.
Euro-peanjournalofoperationalresearch,166(2):547–556,2005.
[74]RenaudLambiotte,VincentD.
Blondel,CristobalddeKerchove,EtienneHuens,ChristophePrieur,ZbigniewSmoreda,andPaulVanDooren.
Geographicaldispersalofmobilecommunicationnetworks.
PhysicaA:StatisticalMechanicsanditsApplications,387(21):5317–5325,2008.
[75]DavidLazer,AlexPentland,LadaAdamic,SinanAral,Albert-LaszloBarabasi,DevonBrewer,NicholasChristakis,NoshirContractor,JamesFowler,MyronGutmann,TonyJebara,GaryKing,MichaelMacy,DebRoy,andMarshallVanAlstyne.
Socialscience:Compu-tationalsocialscience.
Science,323(5915):721–723,2009.
[76]JohnDLeckenbyandShizueKishi.
Howmediadirectorsviewreach/frequencyestimation.
Director,9:9–9,1994.
[77]J.
Leskovec,L.
Backstrom,andJ.
Kleinberg.
Meme-trackingandthedynamicsofthenewscycle.
InProceedingsofthe15thACMSIGKDDinternationalconferenceonKnowledgediscoveryanddatamining,pages497–506.
Citeseer,2009.
[78]JureLeskovec,LadaA.
Adamic,andBernardoA.
Huberman.
Thedynamicsofviralmarketing.
ACMTrans.
Web,1,May2007.
[79]OuyangY.
Li,X.
Reliablesensordeploymentfornetworktrafcsurveillance.
TransportationResearchPartB,45:218–231,2011.
[80]XiaopengLiandYanfengOuyang.
Reliablesensordeploymentfornetworktrafcsurveillance.
TransportationresearchpartB:methodological,45(1):218–231,2011.
[81]JohnDCLittle.
Aggregateadvertisingmodels:Thestateoftheart.
OperationsResearch,27(4):629–667,1979.
[82]JohnDCLittle.
Modelsandmanagers:Theconceptofadecisioncalculus.
Managementscience,50(12supplement):1841–1853,2004.
[83]JohnDCLittleandLeonardMLodish.
Amediaplanningcalculus.
OperationsResearch,17(1):1–35,1969.
[84]HaniIMesak.
Onthegeneralizabilityofadvertisingpulsationmonopolyresultstoanoligopoly.
EuropeanJournalofOperationalResearch,117(3):429–449,1999.
[85]GeorgeEMonahan.
Optimaladvertisingwithstochasticdemand.
ManagementScience,29(1):106–117,1983.
[86]MichaelJNaples.
Effectivefrequency:therelationshipbetweenfrequencyandadvertisingeffectiveness.
AssociationofNationalAdvertisersNewYork,1979.
[87]MarcNerloveandKennethJArrow.
Optimaladvertisingpolicyunderdynamicconditions.
Economica,pages129–142,1962.
[88]M.
E.
J.
Newman.
Thestructureandfunctionofcomplexnetworks.
SIAMReview,45:167–256,2003.
[89]DungNguyen.
Ananalysisofoptimaladvertisingunderuncertainty.
ManagementScience,31(5):622–633,1985.
[90]V.
Nicosia,F.
Bagnoli,andV.
Latora.
Impactofnetworkstructureonamodelofdiffusionandcompetitiveinteraction.
EPL(EurophysicsLetters),94:68009,2011.
[91]Jukka-PekkaOnnelaandFelixReed-Tsochas.
Spontaneousemergenceofsocialinuenceinonlinesystems.
Proc.
Natl.
AcademyofSciences,107(43),2010.
[92]J.
Orwant.
Heterogeneouslearninginthedoppelgangerusermodelingsystem.
UserModelingandUser-AdaptedInteraction,4:107–130,1994.
[93]WeiPan,NadavAharony,andAlexPentland.
Compositesocialnetworkforpredictingmobileappsinstallation.
InProceedingsofthe25thConferenceonArticialIntelligence(AAAI),pages821–827,2011.
[94]G.
Pickard,W.
Pan,I.
Rahwan,M.
Cebrian,R.
Crane,A.
Madan,andA.
Pentland.
Time-criticalsocialmobilization.
Science,334(6055):509–512,2011.
[95]R.
Puzis,M.
D.
Klippel,Y.
Elovici,andS.
Dolev.
Optimizationofnidsplacementforprotectionofintercommunicatingcriticalinfrastructures.
InEuroISI,2007.
[96]RamiPuzis,YuvalElovici,andShlomiDolev.
Findingthemostprominentgroupincomplexnetworks.
AIcommunications,20(4):287–296,2007.
[97]VitorinoRamos,CarlosFernandes,andAgostinhoCRosa.
Societalimplicitmemoryandhisspeedontrackingextremaindynamicenvironmentsusingself-regulatoryswarms.
JournalofSystemsAr-chitecture,FarooqM.
andMenezesR.
(Eds.
),specialissueonNatureInspiredAppliedSystems,Elsevier,Summer,2006.
[98]AmbarGRaoandPeterMiller.
Advertising–salesResponseFunctions.
NewYorkUniversity,GraduateSchoolofBusinessAdministration,1975.
[99]RamCRao.
Advertisingdecisionsinoligopoly:Anindustryequilib-riumanalysis.
OptimalControlApplicationsandMethods,5(4):331–344,1984.
[100]PetriRouvinen.
Diffusionofdigitalmobiletelephony:AredevelopingcountriesdifferentTelecommunicationsPolicy,30(1):46–63,2006.
[101]RolandTRust.
Selectingnetworktelevisionadvertisingschedules.
JournalofBusinessResearch,13(6):483–494,1985.
[102]RolandTRust.
Advertisingmediamodels:Apracticalguide.
Lex-ingtonBooksLexington,MA,1986.
[103]J.
Scott.
SocialNetworkAnalysis:AHandbook.
SagePublications,London,2000.
15[104]SureshPSethi.
Optimalcontrolofthevidale-wolfeadvertisingmodel.
OperationsResearch,21(4):998–1013,1973.
[105]D.
ShahandT.
Zaman.
Rumorsinanetwork:Who'stheculpritArxivpreprintarXiv:0909.
4370,2009.
[106]SrinivasaraghavanSriramandManoharUKalwani.
Optimaladvertis-ingandpromotionbudgetsindynamicmarketswithbrandequityasamediatingvariable.
ManagementScience,53(1):46–60,2007.
[107]R.
Stern,R.
Puzis,andA.
Felner.
Potentialsearch:abounded-costsearchalgorithm.
InAAAI21stInternationalConferenceonAutomatedPlanningandScheduling(ICAPS),2011.
[108]S.
H.
Strogatz.
Exploringcomplexnetworks.
Nature,410:268–276,March2001.
[109]JimSurmanek.
Mediaplanning:apracticalguide,volume49.
NTCBusinessBooksChicago,IL,1996.
[110]GerardJTellis.
Thepriceelasticityofselectivedemand:Ameta-analysisofeconometricmodelsofsales.
JournalofMarketingResearch,pages331–341,1988.
[111]GerardJTellis.
Interpretingadvertisingandpriceelasticities.
JournalofAdvertisingResearch,29(4):40–43,1989.
[112]MLVidaleandHBWolfe.
Anoperations-researchstudyofsalesresponsetoadvertising.
OperationsResearch,5(3):370–381,1957.
[113]YogeshVirkarandAaronClauset.
Power-lawdistributionsinbinnedempiricaldata.
arXivpreprintarXiv:1208.
3524,2012.
[114]S.
WassermanandK.
Faust.
Socialnetworkanalysis:Methodsandapplications.
Cambridge,England:CambridgeUniversityPress.
,1994.
[115]D.
J.
Watts,J.
Peretti,andHarvardBusinessSchool.
Viralmarketingfortherealworld.
2007.
[116]D.
J.
WattsandS.
H.
Strogatz.
Collectivedynamicsof'small-world'networks.
Nature,393(6684):440–442,1998.
[117]D.
R.
WhiteandS.
P.
Borgatti.
Betweennesscentralitymeasuresfordirectedgraphs.
SocialNetworks,16:335–346,1994.
[118]CharlesSolomonLeonidKheitsYehudaJ.
Gur,ShlomoBekhor.
Intercitypersontriptablesfornationwidetransportationplanninginisraelobtainedfrommassivecellphonedata.
TransportationResearchRecord:JournaloftheTransportationResearchBoard,2121:145–151,2009.
[119]S.
H.
Yook,H.
Jeong,andA.
-L.
Barabasi.
Modelingtheinternet'slarge-scaletopology.
ProceedingsoftheNationalAcademyofScience,99(21):13382–13386,Oct.
2002.
[120]ShlomoZilberstein.
Usinganytimealgorithmsinintelligentsystems.
AIMagazine,17(3):73–83,1996.
YanivAltshuleristheCTOandChiefScientistofAthenaWisdom,andaresearcheratMIT(MediaLab).
HereceivedhisBA(withhigh-esthonors)inComputerScienceattheIsraeliInstituteofTechnology,andhisMScandPhD(withhonors)inComputerScienceattheIsraeliInstituteofTechnology.
Altshulerspent3yearsasapost-docatMITattheHumanDynamicsgroup,headedbyProf.
AlexSandyPentland.
Yanivspecializesinbigdata,socialphysics,andnetworkanalysis.
Altshulerhaspublishedover60academicpapersandled15patentapplications.
ErezShmueliisaseniorlectureratthedepart-mentofIndustrialEngineeringatTel-AvivUni-versityandaresearchafliateattheMITMediaLab.
HereceivedhisBAdegree(withhighesthonors)inComputerSciencefromtheOpenUniversityofIsrael,andMScandPhDdegreesinInformationSystemsEngineeringfromBen-GurionUniversityoftheNegev,Israel,underthesupervisionofProf.
YuvalElovici.
Aftercomplet-inghisPhD,Erezspenttwoyearsasapost-doctoralassociateattheMITMediaLab,attheHumanDynamicsgroupheadedbyProf.
AlexSandyPentland.
HismainresearchinterestsincludeBigData,ComplexNetworks,ComputationalSocialScienceandInformationSecurityandPrivacy.
HisprofessionalexperienceincludesveyearsasaprogrammerandateamleaderintheIsraeliAir-ForceandthreeyearsasaprojectmanagerinDeutscheTelekomLaboratoriesatBen-GurionUniversityoftheNegev.
GuyZyskindisagraduatestudentintheHu-manDynamicsgroupattheMITMediaLab.
HeholdsaB.
ScdegreeinElectricalEngineeringandComputerSciencefromTelAvivUniversityandiscurrentlypursuingaM.
SdegreeunderthesupervisionofProf.
Alex"Sandy"Pentland.
BeforejoiningtheMediaLab,Guyhasledthedevelopmentofseveralstart-upsintheBigDataandconsumerspaces.
HisresearchinterestsintersectthestudyofSocialNetworksandBigData,PrivacyandDistributedSystems.
OrenLedermanisagraduatestudentattheMITMediaLab,HumanDynamicsgroup.
HeholdsaB.
Sc.
inComputerScienceandEconomicsfromTelAvivUniversity,Israel.
HisprofessionalexperienceincludesveyearsasdeveloperandteamleaderintheIsraeliarmy,foundingamobilesocialnetworkstartup,fouryearsasdatainfras-tructuresteamleaderinaninuencers-basedmarketingstartup,andaresearchengineerposi-tionattheSingapore-MITAllianceforResearchandTechnology.
Hisresearchinterestsincludebigdataandpersonaldata,groupdynamicsandstudyinginnovationteams.
16NuriaOliver(PhD,MIT2000)iscurrentlytheScienticDirectorandfounderoftheUser,DataandMediaIntelligenceresearchareasinTele-fonicaResearch,workingondataanalytics,ma-chinelearning,usermodelingandHCIinavari-etyofdomains.
Priortothisposition,shewasaresearcheratMicrosoftResearchinRedmond,WAforover7years.
Shehaswrittenover90scienticpapersininternationalconferences,journalsandbookchapters.
Herworkhasbeenwidelyrecognizedbythescienticcommunitywithover7700citations.
Nuriahasover40patentapplicationsandgrantedpatents.
Sheisintheorganizingand/orprogramcommitteeofthetopconferencesinherresearchareas.
Shebelievesinthepoweroftechnologytoempowerandincreasethequalityoflifeofpeople.
Shehasreceivedanumberofawards,includinga10YearTechnicalImpactAward(ACMICMI),aRisingStartAwardbytheWomen'sForumfortheEconomyandSociety(2009),MITsTR100YoungInnovatorsAward(2004)andtheFirstSpanishAwardofEECSgraduates(1994).
SheisseniormemberoftheACM.
Herworkhasbeenwidelyfeaturedonmultiplenewspapers,maga-zines,radioandTVstationsbothinSpainandtheUS.
ShehasbeenfeaturedinELPAISSundaymagazineasoneofafew'femaledirectorsintechnology'(2012),namedRisingTalentbytheWomen'sForumforEconomy&Society(October2009),oneofthe'mostinuentialyoungwomeninSpain'(MujerHoyMagazine,2012),oneof'100leadersofthefuture'byCapitalMagazine(May2009)andoneofthe'GenerationXXI:40SpanishyoungstersthatwillmakenewsintheThirdMillenium'byELPAIS(2000).
ShehasgiventwoTEDxtalksandoneWIREDtalk.
Sheisalsoco-organizingtherstTEDxBarcelonaeventdevotedtoEducation.
Alex'Sandy'PentlanddirectsMITsHumanDynamicsLaboratoryandtheMITMediaLabEntrepreneurshipProgram,co-leadstheWorldEconomicForumBigDataandPersonalDatainitiatives,andisaBoardmemberforNissan,MotorolaMobility,Telefonica,andHarvardBusi-nessReview.
HehaspreviouslyhelpedcreateanddirectMITsMediaLaboratory,theMediaLabAsialaboratoriesattheIndianInstitutesofTech-nology,andStrongHospitalsCenterforFutureHealth.
In2012ForbesnamedSandyoneofthe'sevenmostpowerfuldatascientistsintheworld',alongwithGooglefoundersandtheCTOoftheUnitedStates,andin2013hewontheMcKinseyAwardfromHarvardBusinessReview.
Heisamongthemost-citedcomputationalscientistsintheworld,andapioneerincomputa-tionalsocialscience,organizationalengineering,wearablecomputing(GoogleGlass),imageunderstanding,andmodernbiometrics.
Hismostrecentbookis'SocialPhysics,'publishedbyPenguinPress.