Download presentation
1 -

ChapterVLIDULocationbasedapproachtoIDentifysimilarinterestsbetweenUs


Contents1Pro12lebuilding1002Multi-layerdatarepresentationbasedonuserroutines10421Datarepresentation107211Multi-layerrepresentationofcorrelatedtra-jectories112212Representationoftemporaldata1143Thetraj

claire's Recent Documents

navigation made easy
navigation made easy

e-NAVTORShipping e-Navigationand opportunitiesfor MyOcean Bjrn ge HJLLOProduct / e-NAV ManagerMyOcean UW LisbonJune 20th 2014Norwegian company established in June 2011Business idea is e-Navigationmade

published 0K
by the agreed deadline This is sometimes done through online peer your
by the agreed deadline This is sometimes done through online peer your

11/128Author Hub 10/12may be requested includeIf the article has major structural issues that need to be rectified by significantly reorganizing the textIf more experiments are needed to support the

published 0K
Digital signature
Digital signature

Original on fileDATEApril 25 2019ARCREVIEW CODEV1904251TOChairmanRobbPitts Fulton CountyThe Atlanta Regional Commission ARC has received the following proposal and is initiating a regional review to s

published 1K
DOCUMENT RESUMEED 360 592CG 024 973AUTHORHallard LynnTITLECurriculum A
DOCUMENT RESUMEED 360 592CG 024 973AUTHORHallard LynnTITLECurriculum A

Cartier for Substance Abuse PreventionAlcohol and OtherDrug-Free by the Year 2000esource GuideCURRICULUM-14117-111111111US DEPARTMENT OF EDUCATIONOffice of Educattonal Research and ImprovementEDUCATIO

published 0K
2020-01-20-AVR-Compliance-Hearing-Request.pdf
2020-01-20-AVR-Compliance-Hearing-Request.pdf

D D / Z x0012 x0011 / x0018 D x0004 x0018 / / x0011 x001C / x0004 h K x0011 x001C / x0011 x001C d

published 0K
Key conditions and
Key conditions and

R7UMPOAprASRchallengesSaudi Arabia remains highly dependent on the hydrocarbon sector but a reform acceleration beginning in 2016 had made progress on economic diversix00660069cation es-pecially regar

published 0K
donor names DSUSA and
donor names DSUSA and

Signature ail ctivity pplication 44444444444444444444444444444444444444444444

published 0K
UA is an AAEO employer and educational institution and prohibits ille
UA is an AAEO employer and educational institution and prohibits ille

December 232020 du/nondiscriminationCOVID-19Survey in the Municipality of AnchorageDecember 3-5Brief ReportGabriel Garcia PhD MA MPH1Joy Chavez Mapaye PhD2Rebecca Van Wyck3Katie Cueva ScD MAT MPH4Eli

published 0K
Download Section

Download - The PPT/PDF document "" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.






Document on Subject : "ChapterVLIDULocationbasedapproachtoIDentifysimilarinterestsbetweenUs"— Transcript:

1 ChapterVLIDU-Location-basedapproachtoIDe
ChapterVLIDU-Location-basedapproachtoIDentifysimilarinterestsbetweenUsersinsocialnetworks Contents 1Pro lebuilding.......................1002Multi-layerdatarepresentationbasedonuserroutines1042.1Datarepresentation...................1072.1.1Multi-layerrepresentationofcorrelatedtra-jectories.....................1122.1.2Representationoftemporaldata.......1143Thetrajectorycorrelationalgorithmtoidentifysim-ilarinterestsbetweenusersbasedonuser'sdailyrou-tines.............................1154Sharingroutinesbetweenusers.............1245Conclusion.........................125 Wehaveobservedthatpeopleworkorliveindi erentplacesbuthavetra-jectorycorrelationsintheirdailyroutines.Theusers'dailyroutines,therefore,canbecapturedbymobilesocialapplicationsandsharedinvirtualcommu-nitiesinordertoincreasesocialinteractionsinrealcommunities.Sincewehavenotedthisviabilitytoincreasesocialinteractionsinrealcommunitiesandthelargewidespreadofsmartphonesandsocialnetworks,weproposeamiddlewareofservicesbasedonuser'sdailyroutines,calledLIDU:Location-basedapproachtoIDentifysimilarinterestsbetweenUsers 98ChapterV.LIDU-Location-basedapproachtoIDentifysimilarinterestsbetweenUsersinsocialnetworks insocialnetworks.Thekeyideaistoincreasesocialinteractionsbyrelatingdailyroutinesandpoints

2 ofinterestbasedontrajectoriesofmobileuse
ofinterestbasedontrajectoriesofmobileusers.Forinstance,amobilesocialapplicationjointlywithasocialnetworkhastobeabletoanswerthefollowingquestions:1.WhichofmyfriendsstopinmypreferredbakeryinGrenobleatthesameperiodoftheday?2.Doanyofmyfriendspassnearmyapartmenttogetfromtheirhometotheirwork?3.WhichofmycontactswillbepassingintothecampusoftheUniversityofGrenobleduringtheweek?1Whilethesequestionsareinterestingtoobtaininformationofsimilaritiesbetweenusers'dailyroutines,somescienti cchallengeswereconsideredinthedesigningofourapproach.Thechallengesaremainlyrelatedtotraditionalandnewproblemsinvolvingsocialnetworks,mobilecomputingtechnologiesandspatialdatarepresentation.Wepointoutthesechallengesasfollow:ˆDeterminetherelationsbetweenusersofsocialnetworks.ˆProposeintegratedsoftwarearchitectureaccordingtothecharacteristicsofmobilityscenario,suchaslimitedresources,networkandsensors.ˆDe nethestructureofuserpro lesinordertofacilitatetheassociationbetweentrajectoryandcontextdataofusers.ˆDesignarobustdatamodeltodescribethespatialenvironment,tak-ingintoaccountdi erentlevelsofthespatialinformation.Thedatamodelhelpstheclassi cationofspatialknowledgebasedonpointsofinterest(e.g.,bakery,apartment,campus),spatialrelations(e.g.,nearmyapartment

3 )andgeographicentities(e.g.,Grenoble).&#
)andgeographicentities(e.g.,Grenoble).ˆExtendthedatamodeltorepresenttherelationsbetweenspatialandtemporaldata,whichallowsthecharacterizationofuser'strajectoriesinmultiplecontextinformation.ˆConsidertheaspectsofthequalityofdata(e.g.,sensoreddata)andthesharingofpersonalusers'information(respectingtheprivacyfeatures). 1Theuserde nesthecontactstosharehis/herdailyroutine. 99 ˆExploretheavailableknowledgeinordertoidentifyspatialandcontex-tualsimilaritiesbetweenusers,takingintoaccounttheperformanceandrobustnessoftheapproach.ˆProposeagenericsystemtoprovideadaptableservicesfordi erenttypesofapplications,suchasarecommendationsystem.Withthesechallengesinmindwedecidedtoproposeamiddlewareduetoitsinherentcharacteristics,whichfocusonbridgingthegapbetweenap-plicationsandlow-levelconstructs[169].Amiddlewareisabletoachievetherequirementsofthesepresentedchallengesandprovidemorefeaturesthatfacilitatestheextensionofourapproach,suchasscalability,heterogeneity,dynamicity,adaptability,knowledgemanaging,dataassociation,qualityofserviceandsecurity.Insummary,amiddlewarehelpsdeveloperstocreateap-plicationsthatmakequeriestothemiddlewareservicesandgetresultsbackinanecientway[170].Figure5.1illustratesthearchitectureoverviewofourmiddlewareap-proach.Aswe

4 note,twomaininputdataareacquiredbythemid
note,twomaininputdataareacquiredbythemiddleware,whicharetrajectorydata(jointlywithcontextinformation)andsocialcon-nectionsbetweenusers.Socialconnectiondataaredirectlyprocessedbythedata-modelingalgorithm.ClusteringalgorithmreceivestheGPStrajectoriestodiscoverthebestrepresentativetrajectoryofeachuser.Afterthat,thecorrelationalgorithmidenti esthesimilarpointsofinterestbetweenusers.Finally,thedata-sharingalgorithmisabletoadapttheinformationaccordingtotherequirementsofeachapplication.Formally,ourmiddlewareallowstheexecutionofalgorithmstocapture,store,processandsharesimilarityinformationderivedfromusers'dailyrou-tines.Firstly,weusesmartphonesandtheirsensorstocaptureusers'dailyroutinesandcontextinformation.Secondly,allinformationistransferredandstoredinarelationaldatabaselocatedonaserverapplication,whichisusedasaplug-inonasocialnetworkplatform.Besidesthat,weexploretheca-pabilitiesprovidedbyclusteringalgorithmstoanalyzeusertrajectoriesandextractrelevantinformationfromalargeamountofdata.Finally,weuseanoptimizedtrajectorycorrelationalgorithmtoidentifysimilarroutinesbetweenfriendsinsocialnetworks.Althoughthecoreofourmiddlewareissituatedinthemiddleofthepre-sentedarchitecture,thedataacquisitionprocesshastobewellde nedinorder 100ChapterV.LIDU-Location-basedapp

5 roachtoIDentifysimilarinterestsbetweenUs
roachtoIDentifysimilarinterestsbetweenUsersinsocialnetworks Figure5.1:Architectureoverview.toprovidetherelevantdatatoouralgorithms.Therefore,wepresentthedataacquisitionprocess,calledpro lebuildingentity.Thepro lebuildingcanbedenotedasanalgorithmtoacquiretrajectorydataandtheircontextinfor-mationthroughtheuseofsmartphones.Aftercapturingthepro lebuildingcomponentsendstheacquireddatatothetrajectorycorrelationcomponent,whichisthecoreofourmiddleware.Atthismoment,thealgorithmsintothemiddlewareprocessthedata.ThesetwomainentitiesarepresentedinFigure5.2.Therefore,westartshowingthemainpartsthatcomposethedataacqui-sitionmodule,whichwasadaptedandimplementedtoourapproach.1Pro lebuildingTheuserpro lecanbedeterminedtakingintoaccounttwobasictypesofdatathatareusedforconstructingandenrichingthedatamodel.Thesetwobasictypesarede nedaspersonalandcontextualdata.Personaldatadescribesthemainfeaturesofanentityandthecontextualdatacharacterizesthesituation. 1.Pro lebuilding101 Figure5.2:Maincomponentsofourmiddleware.Anentitycanbeaperson,place,physicalorcomputationalobject.Forex-ample,inapersonaltrackingapplicationformobileusers,thepersonaldatawouldbetheinformationabouttheuser,suchasname,birthday,gender,etc.Ontheotherhand,contextualdatawouldbecomposedofmo

6 vementrecordsthattheuserperformedoverape
vementrecordsthattheuserperformedoveraperiodoftime.Amovementrecordcanin-cludesuchcharacteristicsastheinitialpoint,speed,direction,andtime,aswellasweatherinformation.Wede neanentityasamobileuserusingasmartphoneequippedwithGPS,digitalcameraandInternetconnection(e.g.3GorEdge).Forthecontextualdataorganization,wehavefollowedtheconceptsandrelationsofContextTopontology,introducedbytheauthorsof[7].Figure5.3illustratesthisontology,whereActionhasaContextthatiscomposedbysomeContextElements.ThecontextcanalsodescribethesituationofitselementsthroughthepropertyofdescribeTheSituationOf,whichhasContextisitsoppositeproperty.BasedontheContextTop,wedividethecontextin vemaindimensions:social,spatial,temporal,spatio-temporalandcomputational.Thesocialdi-mensionisrelatedtothefeaturesassociatedwiththeuser,suchasuserpro leandsocialrelationsinasocialnetwork.Thespatialdimensionprovidesthe roPSh  IS in PgSaPnS a #"S r%S I  S S r

7 ;
; S P Sg$S oS Si  S r"S 102ChapterV.LIDU-Location-basedapproachtoIDentifysimilarinterestsbetweenUsersinsocialnetworks Figure5.3:ContextTopontologyconceptsandrelations[7].spatialinformationabouttheenvironmentwheretheactionisdone,forex-ample:geographiccoordinates,postaladdress,etc.Thetemporaldimensioniscomposedbytheinformationabouttime,suchasthedate,thedaysinaweek,etc.Thespatio-temporaldimensionhastheinformationderivedfromthespatialandtemporaldimensions(e.g.,weather).Finally,thecomputa-tionaldimensiono ersthefacilitiesprovidedbytheembeddedsoftwareinthesystem(e.g.,sensors,mobileapplications,etc.).Therefore,thefeaturesthatareusedineachdimensionarede nedbythedeveloperofthecontext-awaresystem.WehaveadoptedthesamedataorganizationpresentedinFigure5.3forde ningthecontextdatageneratedbyourpro lebuildingprocess.Wehavealsode nedathirdtypeofdata,namedbehavioraldata,whichisderivedfromtheassociationamongpersonalandcontextualdata.Weassumethatbeha

8 vioraldataisde nedasauser'sdailyrout
vioraldataisde nedasauser'sdailyroutinethatisgeneratedbasedontheelementsthatcomposeausertrajectoryanditsassociatedcontextualdata.Inotherwords,sincethepersonalandcontextualdataarewellacquiredandassociated,ourapproachallowstheidenti cationofauser'sdailyroutine.Wehavetwowaystoidentifyauser'sdailyroutine,basedonasingletrajectoryorderivedfromasetoftrajectories(e.g.,auserthatgoesfromhometoworkeveryday).Bothwayscanbeexecutedbyfollowingourpro lebuilding 1.Pro lebuilding103 Figure5.4:Thepro lebuildingprocess.process,illustratedonFigure5.4.Theusercanuseamobileapplicationtoregisterasingletrajectorythatdescribeshis/hertrajectorytogofromhometowork,forexample.Aftervisualizingandvalidatingthetrajectorythatrepresentshis/herdailyroutine,theuserpro leiscreatedandthedataissenttothecoreofourmiddleware.Socialconnectionsarealreadyavailablebysomesocialnetworkplatform(e.g.,Facebook,LinkedIn,Twitter)ontheInternet.Therefore,thissingletrajectoryanditscontextualdataareusedtorepresenttheuser'sdailyroutine.Ontheotherhand,thesecondwaytode neauser'sdailyroutineisdiscoveringhis/herbestrepresentativetrajectoryfromasetoftrajectories.Sincetheuserregistersmorethanonetrajectorytorepresentthesamedailymovement,aclusteringalgorithmtechniquecanbeappliedtorecognizetheb

9 estrepresentativetrajectory.Forexample,a
estrepresentativetrajectory.Forexample,ausertookthesimilarpathtogofromhometoworkfor3timesinaperiodof5days.Fortheother2days,thisuserdecidedtochangethepathduetosomeincidentorproblem.Forthisreason,the3similartrajectoriescouldbeusedtorepresentthebestrepresentativetrajectory.Consequently,thisbestrepresentativetrajectoryrepresentstheuser'sdailyroutine.Inourapproach,weprovideamethodtorecognizeauser'sdailyroutinefromoneormultipletrajectories.FollowingthestepspresentedinFigure5.2,thestructuringmoduleveri esifthereisaprevioustrajectoryfortheuser.Ifthereisnotrajectory,itcreatesanewuser'sdailyroutine.Ontheotherhand,ifmultipletrajectoriesarefound,clusteringandaggregationtechniquesareusedtoidentifytheaggregatedtrajectory(abestrepresentativeofuser'sdailyroutine)[171].Aspreviouslymentioned,weapplytheOPTICSalgorithmto aI S rIS g o S i IS hrrn PniSangnrPniS 104ChapterV.LIDU-Location-basedapproachtoIDentifysimilarinterestsbetweenUsersinsocialnetworks classifyusertrajectoriesbasedontheirdailyroutes.Theclusteringandaggregationmoduleprovidesthebestrepresentativetrajectoryforeachuser.Thisaggregatedtrajectoryfromoneuseriscom

10 -paredtootherusersbyapplyingourtrajector
-paredtootherusersbyapplyingourtrajectorycorrelationalgorithm(Section3).Thisapproachenablesgroupsofuserstosharesimilarroutestoincreasegeospatialsocialinteraction.Theuserdailyroutinethenisenrichedwithadditionalinformationabouteachlocationinthedatabase.Thestructur-ingmodulethenexportstheenrichedinformationtoupdatetheuserpro ledatabase.AssumingthatthesedataareavailableontheInternetand,consequently,areconnectedtosomesocialnetworkplatform,thesocialrelationscanbeusedtoenrichthedatabase.Thestructuringmodulerequeststhesocialrelationsforeachuserwhohasregisteredhis/hertrajectoryonthedatabase.Hence,thecomparisonisperformedbasedonthetypeofrelationbetweenusers,forexample:bestfriends,family,colleagues,etc.Inourstudy,weassumedthatthecomparisonoftrajectoriescouldusethisfeatureasa ltertoavoidsecurityproblems,mainlyinvolvingprivacy.Whilethecapabilitiestocaptureasequenceofpositions,toenrichthedatabaseandtodiscoverthebestrepresentativetrajectoryarethestartingpointofmanagingmovement,designingamiddlewarebasedontrajectorydatarequiresastructuralapproach.Afterobtainingthesetrajectories,modelingthembecomesnecessaryforimportantoperations,suchas:i)toindentifypatterns,whichwillbeusedfordecisionmaking(e.g.registeringuserstrajec-torieswithinacityforoptimizingtracofvehicle

11 s);ii)toqueryinformationaboutthemovingob
s);ii)toqueryinformationaboutthemovingobjects(e.g.enrichingtrajectorydatawithcontextinfor-mation);iii)tooptimizeintelligenttransportsystems(e.g.motivatinguserstousecarpoolingalternativesinordertoreducethenumberofvehiclesinurbanregions).2Multi-layerdatarepresentationbasedonuserroutinesThemainmotivationstodesignasuitabledatamodelarerelatedtoprovidinganeasywaytomanipulatetrajectorydata,tousestructuredquerylanguages,tospecifypro lesthroughmovements,tocreateandcomparepro legroups. 2.Multi-layerdatarepresentationbasedonuserroutines105 Inparallel,theidenti cationofthescenarioisasigni cantrequirementtodesignaconceptualdatamodel.Inthisthesis,wetakeintoaccountthescenarioofanemployeethatgoesfromhometoworkandbackeverydaywithinacity,whosetheuser'sdailyroutinecanberepresentedatdi erentabstractionlevels.Inaddition,weconsideradiversityofsemanticdatathatenrichestheknowledgeonthesetrajectories.Forauserdailytrip,wecanobtaininformationaboutpossibleuserinterestsbasedonhis/hermovements.Forexample,whenevertheusergoesfromworktohome,heusuallystopsataspeci cco eeshop.Therefore,theconceptualmodelfortrajectoriesmustbeabletoanalyzeandmanagesimpletrajectories(directtravelsfromorigintodestination)aswellascomplextrajectories(wherethetrajectoryissemanticallyc

12 omposedofseparatesegmentsand/ordi er
omposedofseparatesegmentsand/ordi erentabstractionlevels).Furthermore,thedatamodelmustrelateanytypeofsemanticannotationtotrajectories,suchasattributesofeachtrajectoryandconnectionsbetweenthetrajectoryandanobjectstoredinthedatabase.Often,itisimportanttounderstandthemovementdataatmultipleab-stractionlevelsforpatternrecognitionandanalyzingmovementbehaviorsaswellastodeducetherelationshipsbetweenusersinlocation-basedsocialnet-works.Inordertocreatea exibledatamodelformobilesocialapplicationcontext,weproposeamulti-layerdatarepresentationofmovingobjectsbasedonuserroutines.Aspeci cplaceaswellasasegmentorawholetrajectorycandenotetheseuserroutines.Severalresearchershaveshownaninterestinanalyzingandrepresentingspatio-temporaldata[15][6][14].Thisdataisrelevantinanumberofareassuchassocialinteraction,datamining,medicineandgeographicalinforma-tionsystem.Forinstance,inthecontextofsocialinteraction,wepointedoutsomeapproachesrelatedtocollectingandanalyzingdailytrajectoriesofhumans,addressingissuessuchasdailyroutine,mobility,sport,trips,andsocialnetworks.Inalltheseapproaches,theamountofdataproducedisverylargeandisthereforechallengingtointerpret.Inparallel,theneedforrepresentinginformationaboutPoIontheWebhasemerged[3]inordertomanageandorganizecontext-awareinformation

13 .Interestingissuesincludehowpointsorregi
.Interestingissuesincludehowpointsorregionscanbecorrelatedthroughmulti-layerrepresentation[172]andhowusertrajectoriescouldbeanalyzedintermsoftheirdistancetoanotherone[173]. 106ChapterV.LIDU-Location-basedapproachtoIDentifysimilarinterestsbetweenUsersinsocialnetworks Multi-layerdatarepresentationhasbeenofinterestforalongtimeduetoitsimportanceforspatialdatarepresentation[174][175][176].Inspiteofthelargenumberofissuesaboutmulti-layerdatarepresentation,thereisalackofmulti-layerrepresentationtechniquesformovingobjecttrajectories.In[177],theauthorspresentadesignformulti-layerspatialobjectsinwhichbothspatialobjectsandtheverticesoftheircomponentgeometryarelabeledwithlevelpriorityvalues.Althoughthedatamodelsupportingqueriesatdif-ferentabstractionlevelsisveryinteresting,itisnotintendedforrepresentingtrajectoriesandnoteasilyextendableforthiscontext.In[178],theauthorspresentaninterestingRule-basedLocationPredictionmethod(RLP),toguesstheuser'sfuturelocationforlocation-basedservices.However,theydonotconsiderthepartialcontainmentrelationshipbetweenspatialregionsatdi erentspatiallevels.In[179]and[180],theauthorsin-troduceapproachestoconsidertrajectorypatternsbetweendi erentspatiallevelsaswellastherelationamonguser,locationandtrajectory.Inparticular,GeoLife[179]i

14 sasocialnetworkingservicewhichincreasess
sasocialnetworkingservicewhichincreasessocialconnectivityamonguserstakingmultiplegeospatialscalesintoaccountwhiletheworkde-scribedby[180]focusesonRegionsofInterest(ROI)asopposedtomultipleabstractionlevels.InthisthesisweextendthePoIdatamodelproposedbyW3Cworkinggroupandpresentamulti-layerdatarepresentationofcorre-latedtrajectories,takingintoaccountthePoIatmultipleabstractionlevels.AsintroducedinChapterIV,PoIiscomposedofanynumberofthefol-lowingentities:ˆlabel:isahumanexplicitlabeltonamePoI.Thisentityisimportanttoidentifyaspeci cplace,whichcanbeusedtosupportthede nitionsoflabelsinthedi erentlevelsofourdatamodel.ˆdescription:ahumanexplicitdescriptionaboutthePoI.ˆcategory:thisentityclassi esPoIintoacategory.Forexample,itcanbeaprimaryattribute(e.g.,museum,bar,restaurant),apopularityranking,orasecurityrating.ˆtime:timeisconsideredthemostcommoncontextinformation,whichisgenerallyrepresentedbythetimeinstantthatthelocationwasac-quired.Timeisalsousedtoestimatethedurationofanobjectataplace[24]. 2.Multi-layerdatarepresentationbasedonuserroutines107 ˆlink:thisentityisagenericmannertorepresentarelationshipfromaPoItoanotherPoI,orfromawebresourcetoaPoI,bothbasedontheRFC4087technique(point-to-pointlink).ˆmetadata:inthisentity,we

15 caninsertformalmetadatatothePoI(byrefere
caninsertformalmetadatatothePoI(byreference,forexample).Therefore,wehaveusedthisde nitiontoconstructourdatarepresenta-tion,whichispresentedinthenextsection.2.1DatarepresentationInourwork,weassumethattheinterestsofauserforaspeci cplace,segmentortrajectorycanrepresentauserroutine.Forinstance,auserlikestoeatattherestaurantXeveryday,wherethisrestaurantisapointofinterest.Inthesameway,auserpreferstotakeaspeci cstreet(segment)orasetofdi erentsegments(trajectory)togofromhometowork.Alongthisline,auserroutinecanbede nedfollowingamulti-layerrepresentation(seeFigure5.5),wherenrepresenttheidenti erofeachelementoftheroutine,andthelinksaretherelationsbetweentheseelementsatdi erentabstractionlevels.TakingintoaccounttherepresentationpresentedinFigure5.5,weclassifyuserroutinesasTrajectoryofInterest(ToI),SegmentofInterest(SoI)andPoIatdi erentabstractionlevels.Therefore,wede nethisspatialinformationtobeamulti-layerdatarepresentationinordertosupportthedescriptionoftheuser'sdailyroutine.AccordingtoFigure5.5,auserroutineispresentedbasedonitslayer.Forinstance,thelastlayer(Layer3)canberepresentedbythenameofthelocationaccordingtotheGPScoordinate(e.g.bakery'sname,housenumber,etc.),basedonthePoIdatamodelproposedbyW3Cworkinggroup(withthesameentiti

16 es).Nevertheless,weinsertedtheentitycall
es).Nevertheless,weinsertedtheentitycalleduser idtoidentifytheownerofthePoI.Inparallel,wereusedandadaptedthePoIdatamodeltode netheentitiesandvaluesofSoIandToI.Followingourmulti-layerdatarepresentationandthereferencemodelofW3C,theLayer2isde nedastheSegmentsofinterest(SoI)thatcomposetheusertrajectory,whereeachSoIiscomposedofanynumberofthefollowingentities:ˆuser ID:isusedtoidentifytheownerofSoI. 108ChapterV.LIDU-Location-basedapproachtoIDentifysimilarinterestsbetweenUsersinsocialnetworks Figure5.5:Ourmulti-layerdatarepresentation.ˆlabel:isahumanexplicitlabeltonameSoI,whichcanbegeneratedbyusingthelabelsofPoI(e.g.,fromWorktoBakery).ˆdescription:ahumanexplicitdescriptionabouttheSoI.ˆcategory:theclassi cationofSoIintoacategory.Similartothecate-goryofPoI,itcanbeaprimaryattribute(e.g.,street,avenue,highway),apopularityranking,orasecurityproperty.ˆtime:forthisentity,wecanhavethetimeintervalthatthemovingobjectstayedintoSoI,basedontheinitialand naltimeinstants.Thesetimeinstantsarederivedfromthetimeinstantsofthecorrespondinginitialand nalPoI'softhesegment.ˆlink:similartothePoI,thisentityisagenericmannertorepresentarelationshipfromaSoItoanotherSoI,wherethelastPoIofthesegmenthasalinkwiththeinitialPoIofthenextsegment. So

17 I(n+1)ToI PoI n PoI(n+1) 2.Multi-layerda
I(n+1)ToI PoI n PoI(n+1) 2.Multi-layerdatarepresentationbasedonuserroutines109 ˆmetadata:ittheuseofaformalmetadatatoSoI(byreference,forexample).Finally,ToIintheLayer1couldberepresentedbyawholeusertrajectory(e.g.togofromhometowork).Therefore,weidentifythefollowingentitiesthatcomposeeachToI:ˆuser ID:isusedtoidentifytheownerofToI.ˆlabel:isahumanexplicitlabeltonameToI,whichcanbealsogen-eratedbyusingthelabelsofPoI(e.g.,fromWorktoHome)orbythelabelsofSoI(e.g.,fromstreetXtoavenueY).ˆdescription:ahumanexplicitdescriptionabouttheToI.ˆcategory:theclassi cationofToIintoacategory.SimilartothecategoryofSoI,itcanbeaprimaryattribute(e.g.,nameoftheregionthatthewholetrajectorywasregistered),apopularityranking,orasecurityproperty.Themaincharacteristicforthesecuritypropertyisrelatedtotheaccesscontrolpolicesforausertrajectory.Basedontheleveloftherelationshipwithanotheruser,theusercancontrolthesharingofthewholetrajectory(e.g.,Public,List of Group Access(speci cgroupoffriendsinmysocialnetwork),PrivateorListofusers).Whilethispropertycanbede nedbytheuserinToI,itcanbealsode nedinthesecuritypropertiesofSoIandPoI.ˆtime:thetimeintervalthatthemovingobjectstayedintoToI,basedontheinitialand naltimeinstants.SimilartothetimeentityofSoI,th

18 esetimeinstantsarederivedfromthetimeinst
esetimeinstantsarederivedfromthetimeinstantsofthecorre-spondinginitialand nalPoI'softhetrajectory.Inaddition,withthisinformationwecanidentifytheperiodofthedayandthedaysoftheweek,forexample.ˆmetadata:ittheuseofaformalmetadatatoToI(byreference,forexample).Basedonthisstructure,auserroutineispresentedasageneralinterestaccordingtotheabstractionleveloftheuser/system.Besidesthat,aToIisdirectlyrelatedtoasetofSoI'sand/orPoI'satlowlevels.Tobetter 110ChapterV.LIDU-Location-basedapproachtoIDentifysimilarinterestsbetweenUsersinsocialnetworks understandthisrelation,weuseatreestructuretoshowtherelationbetweeneachinformationaccordingtothemulti-layerdatarepresentation.Basedontheillustrateddatarepresentation,wedesignourmulti-layerdatamodelfortrajectories,takingintoaccountdi erentabstractionlevelsofuserroutines.Inthefollowingweprovidethebasicde nitionstosupportourdiscussion.1.Trajectory(T)isde nedasasetofconsecutivepointscapturedthroughaGPSofonetripperformedbyauser.Eachlocation(L)iscomposedofasetofinformation(latitude,longitude,altitude,direction,timestampforeachregisteredpoint(tL)andanapproximatespeedprovidedbytheGPS).T=fL1,L2,L3,:::,Lng,thetimeintervalbetweentwopointsiscomputedbythesubtractionoftL(k+1)�tL(k),where(1kn).Thistemporalinformation

19 alsoallowstherecognitionofpauseinstants,
alsoallowstherecognitionofpauseinstants,accordingtotheproposalof[24].Althoughthepointsarecharacter-izedbylatitude,longitudeandaltitude,wefocusonpointsin2Dspace(latitudeandlongitude)torepresentthepositionofeachuser.2.UserRoutine(UR)isde nedasahumanconstructtorepresentaroutineofauserbasedonhis/herinterest.URtypicallydenotesauserinterest,whereausercanidentifyanentiretrajectory,segmentofroute(e.g.streetname)orplace(e.g.,bakeryX),accordingtothelayerspresentedinFigure5.5,typicallyrepresentedbynameandcharacterizedbytype,whichmaybeusedasareferencepointoratargetinalocationbasedservicerequest(e.g.,routedestination).3.SetofUR(SUR)isde nedasthesetofuserroutinesbasedontheabstractionlevelofmulti-layerrepresentation.Theuserroutineofeachabstractionlevelisde nedaccordingtoitsidenti er(ur),suchthattraj,segandpoirepresentstheToI,SoIandPoIrespectively.There-fore,SURisformedbya nitesetandsubsetofuserroutinesindif-ferentabstractionlevels,e.g.SUR=furtrajfur(seg;1);ur(seg;2);ur(seg;n)g,ursegfur(poi;1);ur(poi;2);ur(poi;n)g,:::,ur(s�1)fur(s;1);ur(s;2);ur(s;n)gg,wheresrepresentstheabstractionlevel.Forinstance,thesettorepresentausertrajectoryinthecampusofJosephFourierUniversityis 2.Multi-layerdatarepresentationbasedonuserroutines111 SURtraj=fChemistryStreetfG

20 renobleInformaticsLaboratory;CERMAVLabor
renobleInformaticsLaboratory;CERMAVLaboratoryg;PiscineStreetfENSIMAGLaboratoryg;LibraryStreetfCentralLibrary;Mathe�maticsLaboratoryggwheretrajcanberepresentedbytheusertrajectoryintheLayer1,Chem-istryStreet,PiscineStreetandPiscineStreetareroadsegmentsintheLayer2,andGrenobleInformaticsLaboratory,CERMAVLaboratory,ENSIMAGLaboratory,CentralLibraryandMathematicsLaboratoryarelocalplacesintheLayer3.Theintentiontodesignaconceptualmodelistoo erbasicproceduresinordertosupportdesignersinthedevelopmentofmobilesocialapplications.Ausualfeatureinthespatialmulti-layerdatamodelistheuserroutinecor-respondingtoagivenabstractionlevel(trajectory,roadsegmentandlocalplace).Whenweconsiderthatagraphofuserinterestsisatree,wecansaythatauserinterestisassociatedwithurindi erentabstractionlevels,whichallowstoindicatethatauserinterestbelongstotheabstractionlevels(traj,segandpoi)associatedwithur.Sincethemulti-layerdatarepresentationispresented,wetakeintoaccounttheorganizationofobjectsforade nedabstractionlevel.Consequently,alowabstractionlevelo ersthesetofPoI'stodescribeausertrajectoryatthehighestabstractionlevel.Weobservethatforalluserroutineshowninthedatarepresentation,wemayhaveaspeci cURavailableateachabstractionlevel(s),suchthatL2urpoi.Thisrepresentationo

21 ersaproceduretounderstandthesetofabstrac
ersaproceduretounderstandthesetofabstractionlevels.Finally,sincetwousersAandBhavearelationinthesocialnetwork,ourdatamodelallowstheidenti cationofsimilaruserroutinesbetweenthem,takingintoaccountthedi erentsituations,presentedinFigure5.6.Wecon-sidertherepresentationofthreemainsituationsofsocialinteractionbetweenusers.Forthe rstsituation(Figure5.6(a)),weobservethattwousershaveapointofinteractionatthecrossingoftwoUR's(e.g.roadsegment).AssumingthatauserApassesinaspeci cregion(e.g.,atcampusofUniversityofGrenoble)andtheuserBalsopassesatthiscampus,wecannotarmthatbothusersaresharingalocationLinurpoi.However,ourdatamodelprovides 112ChapterV.LIDU-Location-basedapproachtoIDentifysimilarinterestsbetweenUsersinsocialnetworks (a)Crossing. (b)SharingaUI. (c)Nearing.Figure5.6:Threemainrepresentationsofsituationsthatweconsiderassimilaritiesbetweentwousers.amannertoidentifythiscrossingindi erentabstractionlevels.Sinceweidentifycommonregionsbetweenbothusers,wecanidentifysimilarsegmentsandpointsofinterest,allowingtheidenti cationofsimilarroutinesindi erentabstractionlevels.InFigure5.6(b)thepointofinteractioncouldbethecompletesetofPoI(e.g.allroadsegmentorapartofit).Forthisexample,whenweidentifythatbothusersaresharingastreet,itisnotevidentthatt

22 heyaresharingthesamepartofthissegment.Ho
heyaresharingthesamepartofthissegment.However,whilethesimilarsegmentisidenti ed,ouralgorithmsverifyifthelocationsrepresentedinthePoIlayercorrespondstothesamepartofthesharedsegment.Finally,inFigure5.6(c),themostimportantinformationistheproximitybetweenusers.Hence,thisproximitycanbedeterminedaccordingtoeachlayerinourmodel.Theusercouldde nethisproximity.Consequently,theuserscanconsiderapossiblesocialinteractionduetotheproximityoftheirtrajectories,segmentsorpointsofinterest.2.1.1Multi-layerrepresentationofcorrelatedtrajectoriesAsoneorasetofuserinterestsmaydescribeauserroutine,weneedtoconsidereveryinformationofeachabstractionlevel(ToI,SoIandPoI).Wethende neausertrajectoryasasequenceofUR's,wherethesetofsegmentscrossesbetweendi erentabstractionlevelsintherequiredorder.Thefollow-ingexamplepresentsamulti-layerrepresentationinordertoillustrateourapproach.ˆsetToI=furtrajg 2.Multi-layerdatarepresentationbasedonuserroutines113 ˆsetSoI=fur(seg;1);ur(seg;2);ur(seg;3)gˆsetPoI=fur(poi;1);ur(poi;2);ur(poi;3);ur(poi;4);ur(poi;5);ur(poi;6);ur(poi;7)gForinstance,wecanconstructthefollowingsetsofUR(SUR):ˆSUR1=furtrajfur(seg;1)fur(poi;1);ur(poi;2)gggˆSUR2=furtrajfur(seg;2)fur(poi;3);ur(poi;4);ur(poi;5)gggˆSUR3=furtrajfur(seg;3)fur(p

23 oi;6);ur(poi;7)gggTheFigure5.7illustrate
oi;6);ur(poi;7)gggTheFigure5.7illustratesthesesetsofUR'srelatedtoeachabstractionlevel.Inthenextde nition,theuserroutinedescriptor(D)containsthesequenceofthedetermineduserroutines.Forinstance,wedeterminetwodi erenttrajectorydescriptorsforuser1(D1)anduser2(D2):ˆD1=ur(seg;1);ur(seg;2);ur(poi;7)&#x-278;ˆD2=ur(poi;1);ur(poi;3);ur(poi;5)&#x-278;WenotethatthedescriptorscanbecomposedbyUR'satdi erentab-stractionlevelsduetomultiplelocationnames,whichcanbeobtainedfromreversegeocodingservices.Therefore,ourdatarepresentationisalsoableto ndasimilarityalthoughtheseUR'sareatdi erentabstractionlevels.Theconceptofmulti-layerrepresentationisanimportantsteptounderstandtherelationsandsimilaritiesbetweenUR's,groupedindi erentuserdescriptors.Forinstance,ifweconsiderD1,theuserdescribesatrajectoryfromadepar-tureurseg(atthesecondabstractionlevel)toadestinationinaurpoi(atthethirdabstractionlevel).IncaseofD2,theuserdescribeshis/herroutineatthesamelevel.Amulti-layerdatarepresentationshouldbeabletoidentifytheabstractionlevelofeachUR.Thisdatarepresentationbecomesanimportantelementforprovidingtheaccurateinformationtoidentifythesimilaritiesbetweenuserroutines.IfweobservetheprevioustrajectorydescriptorsandthethreesituationspresentedinFigure5.6,weseesomec

24 hallengestodevelopadatamodelatdi ere
hallengestodevelopadatamodelatdi erentabstractionlevels.Forinstance,ifweobserveD1andD2,weobservethatthe rstuserispassinginur(seg;2)(atthesecondlevel)andtheotheruserispassinginur(poi;3)(atthethirdabstractionlevel).Therefore,ourapproachallowstheidenti cationofsimilarroutinesbetweenuserswhoaresharingUR'sindi erentabstractionlevels. 114ChapterV.LIDU-Location-basedapproachtoIDentifysimilarinterestsbetweenUsersinsocialnetworks Figure5.7:Exampleofmulti-layerdatarepresentation.2.1.2RepresentationoftemporaldataWhiletheclusteringalgorithmprocessesthespatialinformationinordertoidentifythebestrepresentativetrajectoryforeachuser,thetemporalinfor-mationbecomesrelevantcontextualinformationtoenrichtheservicesthatareprovidedbyourmiddleware.Hence,wedesignedadatarepresentationoftemporalinformation,whichisdetailedasfollows.Ourapproachfollowsthetemporalrepresentationpresentedin[148],wherethetimeisprocessedafteridentifyingthespatialsimilarities.Makinguseofthebestrepresentativetrajectories,weobtainmultipleinformationoftimeforeachpositionintheuser'strajectory.Figure5.8illustratesanexampleofabestrepresentativetrajectorywithintermediarylocations.Assumingthatthisbestrepresentativetrajectorywasobtainedbyadatasetof10trajectoriesofausertogoeverydayfromhometowork.Conse

25 quently,wehave10workingdaysforthisexampl
quently,wehave10workingdaysforthisexample.Theclusteringalgorithmthendiscoversthattheuserrecorded7similartrajectories,bypassingatthesamestreetsandneartospeci clocations.Intuitively,wenotethatthisuserreg-istereddi erenttimeinstantsbylocation(illustratedbythepointsinthetrajectory).Wecanseethesedi erenttimeinstantsinTable5.1.InTable5.1,wemaydeducethattheuserhavetraveledforthreedi erenttrajectoriesinthreeworkingdaystogofromhometowork.ThesedaysareDay2,Day6andDay10.Incontrast,wehaveseventrajectoriesthatwererecognizedtoconstructthebestrepresentativetrajectory.Giventhe 3.Thetrajectorycorrelationalgorithmtoidentifysimilarinterestsbetweenusersbasedonuser'sdailyroutines115 Figure5.8:Exampleofabestrepresentativetrajectorywithmultiplelocationsbetweenadeparture(Home)andadestination(Work). Locations Day1 Day3 Day4 Day5 Day7 Day8 Day9 Home 08:00 08:10 08:05 08:07 08:15 08:12 08:17 Bakery 08:07 08:18 08:12 08:15 08:23 08:20 08:25 Supermarket 08:15 08:26 08:20 08:23 08:30 08:28 08:32 Restaurant 08:25 08:36 08:29 08:31 08:37 08:35 08:39 Postmail 08:32 08:43 08:36 08:38 08:42 08:41 08:45 Work 08:40 08:50 08:43 08:45 08:50 08:47 08:52 Table5.1:Timeinstantsbylocationfromabestrepresentativetrajec-toryofauser.Supermarketasthelocation,weseethattheuserpassedcloseto

26 itat08:15inthe rstday,at08:26intheth
itat08:15inthe rstday,at08:26inthethirddayandatdi erenttimeinstantsintheother5days.Takingintoaccountthisexample,wedesignedourrepresentationoftem-poraldata,wherethekeyideaistostoreallthetimeinstantsbylocationandrepresenttheminatimeinterval.Thetimeintervalspeci esallthetimeinstantsthattheuserpassedclosetoeachspeci clocation.Finally,thisdatacanbeusedtoenrichtheinformationthatwillbeprovidedbyourmiddleware.3Thetrajectorycorrelationalgorithmtoiden-tifysimilarinterestsbetweenusersbasedonuser'sdailyroutinesTakingintoaccounttheideatoanalyzeuser'sdailyroutinesinordertoin-creasethenumberofsocialinteractionsbetweenusers,weproposeanopti- 116ChapterV.LIDU-Location-basedapproachtoIDentifysimilarinterestsbetweenUsersinsocialnetworks mizedalgorithmbasedonMinimumBoundingRectangles(MBR)[181]andtheHausdor distance[182].TheHausdor distanceisoftenusedtodeterminethesimilarityoftwoshapes[183]andtomeasureerrorsforapproximatingasurfaceingeneratingatriangularmesh[184].Inourapproach,weareinterestedtouseHaus-dor distancecomputationintwodi erentcases.Basically,the rstcaseisappliedwhenthealgorithm ndsacorrelatedareabetweentwoMBR's.ItusesHausdor distancetocomputethedistancebetweenthepointsthatareinthecorrelatedarea.Ontheotherhand,ifthereisnoc

27 orrelatedarea,theHausdor distancecom
orrelatedarea,theHausdor distancecomputationisusedtocomputethedistanceofnearpointsbetweentwoMBR's.WhenthedistanceoftwoMBR'sisfound,thealgorithmallowstheexpansionofbothMBR'sinorderto ndoneormorepointsofsocialinteractions,takingintoaccountathreshold(Dmax)fortheexpansion.Firstly,weidentifyfourextremepointsofeachtrajectory(thenorthern-most,thesouthernmost,thewesternmostandtheeasternmost).Withthesepoints,wecreatetheMBRfortheusers'trajectories.Figures5.9illustratestheMBRforaspeci ctrajectory. Figure5.9:AnexampleofMBR.Forinstance,weconsidertwousersAandBandtheexistenceofMBR'sfortheirrespectivetrajectories.ThefourpointstorepresenttherectangleoftheuserAare:(Latmax(A);Lonmin(A));(Latmax(A);Lonmax(A));(Latmin(A);Lonmin(A));(Latmin(A);Lonmax(A)):ThepointsfortheuserBare:(Latmax(B);Lonmin(B));(Latmax(B);Lonmax(B));(Latmin(B);Lonmin(B));(Latmin(B);Lonmax(B)): (Lat, LonX) , LonY) , LonX) , LonY) 3.Thetrajectorycorrelationalgorithmtoidentifysimilarinterestsbetweenusersbasedonuser'sdailyroutines117 Furthermore,weexecutethetrajectorycorrelationprocessaccordingtheal-gorithmasfollows. Algorithm2Mainalgorithm. Input:twotrajectoriesofusersAandBwiththepointscontainingtheirrespectivecoordinates.Comment:Itisveri edifthetwoMBR'sdoesnothaveacorrelatedarea.if(Latmax(

28 A)Latmin(B))or(Latmax(B)Latmin(A))or(Lon
A)Latmin(B))or(Latmax(B)Latmin(A))or(Lonmax(A)Lonmin(B))or(Lonmax(B)Lonmin(A))thenExecuteHausDistofMBR(A)andMBR(B);ifHausDistDmaxthenExpandMBRs;Restartmainalgorithm;elseThereisnocorrelatedarea;Stopmainalgorithm;endifendifComment:Otherwise,weselectthecorrelatedareaandexecutetheHaus-Distalgorithm.Selectcorrelatedarea(Alg.3);ExecuteHausDist(Alg.4);Output:ThepointsinthecorrelatedareaandthedistancesbetweenthepointsofAinrelationtothepointsofB. Aswecanobserveinthemainalgorithm,whenthereisnocorrelationbe-tweentwoMBR's,weexecuteanalgorithmtocomputetheHausdor distancebetweentwoMBR's.ThemainreasontocarryoutthisalgorithmisrelatedtotheprobleminvolvingextremepointsintheMBRfaces.Forexample,wehaveapointintherightfaceoftheMBR(A)andanotherpointintheleftfaceoftheMBR(B).AlthoughtheMBR(A)isclosetotheleftfaceofMBR(B),theremightbenointersection,aspresentedinFigure5.10.Then,wemight 118ChapterV.LIDU-Location-basedapproachtoIDentifysimilarinterestsbetweenUsersinsocialnetworks haveaproblem,becausetwonearpointsarenotpresentinthecorrelatedarea. Figure5.10:MBRExpansionforthenon-intersectionproblem.Tosolvethisproblem,weproposeaMBRexpansionalgorithm,whichcomputestheHausdor distanceoftwoMBR'sinordertoverifyiftheex-pansionispossibleornotaccordingtothethresholdDmax.TheHausdor

29 distancefromtheMBR(A)totheMBR(B)canbedet
distancefromtheMBR(A)totheMBR(B)canbedeterminedbyexploitingthecharacteristicforeachMBRarea,therehastobeatleastoneobjectthattouchesit.Therefore,weidentifytheareainMBR(A)closesttoafaceinMBR(B).Afterthat,thealgorithmcomputestheHausdor distance(Haus-Dist)ofthesetwofacesandcomparetheresultwithDmax.IfHausDistislessthanDmax,thenbothMBR'sexpandstheirrelatedareasfromthecurrentdistancetotheresultofDmax.Figure5.10showstheMBRexpansionprocessforthenointersectionproblem.Ontheotherhand,ifthereisanintersectionofMBR's,thealgorithm3isexecutedinordertodeterminethecorrelatedarea.SincethecorrelatedareaofMBR'sisfound,themainalgorithmexecutestheHausdor distancecomputationofthepoints.AssumingthataandbarepointsofsetsAandBrespectivelyandthattheyareinthecorrelatedarea,thentheAlgorithm4isexecuted. Non-intersection problem MBR(A) MBR(B) Expanded MBR(A) HausDist Dmax Expanded MBR(B) Intersection Intersection 3.Thetrajectorycorrelationalgorithmtoidentifysimilarinterestsbetweenusersbasedonuser'sdailyroutines119 Algorithm3Selectionprocess Input:Latmin(A),Latmax(A),Latmin(B),Latmax(B)ifLatmax(A)�Latmax(B)thenSelectLatmax(B)elseSelectLatmax(A)endififLatmin(A)�Latmin(B)thenSelectLatmin(A)elseSelectLatmin(B)endififLonmax(A)�Lonmax(B)thenSelectLonmax(B)elseS

30 electLonmax(A)endififLonmin(A)�Lo
electLonmax(A)endififLonmin(A)�Lonmin(B)thenSelectLonmin(A)elseSelectLonmax(B)endifOutput:correlatedarea 120ChapterV.LIDU-Location-basedapproachtoIDentifysimilarinterestsbetweenUsersinsocialnetworks Algorithm4Hausdor distancealgorithm Input:pointsoftrajectoriesA(aisuchasi=1ton)andB(bjsuchasj=1tom),wherenandmarethetotalofpointsinthetrajectoriesAandBrespectively.HausDist=0forallpointaiofAdoshortest=Inf;forallpointbjofBdodistanceij=distance(ai,bj)ifdistanceijshortestthenshortest=distanceijendifendforifshortest&#x]TJ/;༕ ;.9;Ւ ;&#xTf 1;.99; 0 ;&#xTd [;HausDistthenHausDist=shortestendifendforOutput:theshortestdistanceofapointinthetrajectoryAandanotherpointinthetrajectoryB. 3.Thetrajectorycorrelationalgorithmtoidentifysimilarinterestsbetweenusersbasedonuser'sdailyroutines121 Whilethesimilaruserroutinesareidenti edbetweentwobestrepresenta-tivetrajectories,thealgorithmstartsthecomparisonbetweentemporalinfor-mation.Tocomparethetemporalsimilaritiesbetweenusers,weconsiderallthesimilarlocationsidenti ed.WehaveadoptedtheParzen-windowmethod[185]toidentifytemporalsimilaritiesbylocation.Parzen-windowhasbeenusedinalargenumberofresearchareas,suchaspatternrecognition,dataclassi cation,imageprocessingandtracking.Wedecidedtousethe

31 Parzenwindowduetothewellrepresentationof
Parzenwindowduetothewellrepresentationofeachtimeinstantatthetimeinter-val,wherethedensityofthepointscanbeeasilyrecognizedandvisualizedinthegraph.Byde nition,theParzen-windowisadensity-basedestimationthatcon-sidersthedata-interpolationtechnique[186].Assumingthatwehavearan-domvariable(x),thenthistechniquecomputestheprobabilitydensityfunc-tion(PDF)inwhichtherandomvariablewasderived.Insummary,itsuper-poseskernelfunctionsateachobservation(xi).Hence,thePDF(f(x))oftheParzen-windowiscomputedbyf(x)=1 nnXn=11 hdimnKx�xi hn;(5.1)whereK()isthekernelfunction,dimisthedimensionalspaceandhnisthewindowwidth.Basedonthisequation,weareabletocomputethevalueoff(x)atacertainlocation(point).Alongthisline,wecandetermineawindowfunctionatxandde nethetotalofobservationsxithatareclosetothewindow.Forourapproach,wedeterminedtheGaussianPDFasthekernelfunc-tionforParzen-windowdensitycomputation.Thus,thePDFf(x)withtheGaussianfunctionbecomesf(x)=8�&#x]TJ ;� -2;.51; Td;&#x [00;:1 nnPk=11 (hp 2)dime�1 2(x�xk h)2iftbxte0otherwise;(5.2)suchastbistheinitialtimeinstantandteisthe naltimeinstantwithinthetimeintervalforeachlocation.Asweareanalyzingapointincomparisontoanotherpointsinthetimeinterval,thevalueofdim=1.Animportantelementrela

32 tedtotheuseofParzen-windowisthevalueofth
tedtotheuseofParzen-windowisthevalueofthewindowsize(h).Accordingto[187]and[188],whentheGaussiankernelisbeingused,the 122ChapterV.LIDU-Location-basedapproachtoIDentifysimilarinterestsbetweenUsersinsocialnetworks optimalvalueofhisde nedbyh=45 3n1 5;(5.3)wherenisthenumberoftimeinstantsinthetimeintervalandisthestandarddeviationofthesamples.Therefore,wecanobtainthefrequencythattheuserisneartoacertainlocation.Sincewehaveidenti edasimilarroutinesbetweentwousers,wecancomparethetemporalgraphstoknowtheprobabilityofrendezvousbetweenthematacertainperiodoftime.Takingintoaccounttheexampleofthesupermarket(Table5.1)forauserA,weconstructatimeintervalbetween08:15and08:32,withthetimeinstants[08:15,08:20,08:23,08:26,08:28,08:30,08:32].Then,thePDFoftheParzenfunctionthengeneratesthegraphpresentedinFigure5.11inordertorepresentallthetimeinstantsthatuserApassedneartosupermarketinthesesevendays. Figure5.11:TimeinstantsthatuserAhavepassedneartosupermarketinthe7days.Asweobserve,thegraphshowstheprobabilityofeachtimeinstantthatuserAwasneartosupermarketintheinterval.Next,weassumethatanother 3.Thetrajectorycorrelationalgorithmtoidentifysimilarinterestsbetweenusersbasedonuser'sdailyroutines123 userB(whoisfriendofA)havealsopassedneartothesamesupermarketinother

33 tendays.GivenatimeintervalofuserBbetween
tendays.GivenatimeintervalofuserBbetween08:10and08:50,withthetimeinstants[08:10,08:15,08:16,08:16,08:20,08:21,08:30,08:40,08:42,08:50],thePDFoftheParzenfunctiongeneratesthegraphpresentedinFigure5.12. Figure5.12:TimeinstantsthatuserBhavepassedneartosupermarketinthe10days.Intuitively,weobservethatthesegraphscanrepresentalltimeinstantsinwhichbothusershavepassedneartoeachlocation.SincewediscoversimilarroutinesbetweenusersAandB,wecanusethesegraphstoestimatetherendezvousbetweenthem,consideringthetemporalsimilarity.Onewaytocomparethesegraphsisthroughthesuperposition,byobservingthecommonareas.Anothermanneristocomputetheprobabilityfromthehighestvalueoftimeinstanttothelowestvalue. 124ChapterV.LIDU-Location-basedapproachtoIDentifysimilarinterestsbetweenUsersinsocialnetworks 4SharingroutinesbetweenusersAwell-knownsolutionofWebapplicationsthatinvolvessharingandestima-tionofuserinterestsiscalledrecommendationsystem.Ingeneral,recom-mendationsystemsareclassi edintwogroups,whicharecontent-basedandcollaborative lteringsystems[165].Intermsofcontent-basedsystems,arecommendationisperformedbasedontheuserpreferencesinrelationtoaspeci ccontent.Forexample,ifauserpreferstolistencountrymusicthantheothergenres,thesystemrecommendsnewsongshavingthe"country"genreastheprefe

34 renceforthatuser.Ontheotherhand,collabor
renceforthatuser.Ontheotherhand,collaborative lteringsys-temsrecommendsomeinformationbasedonsimilarfeaturesofusersand/ordata.Thiskindofsystemiscommonlyusedtorecommendinformationthatispreferredbyagroupofsimilarusers.Takingintoaccountthecharacteristicsofourapproach,wehaveconsideredthecollaborative lteringsystemasthebestmethodtosharetheroutinesbetweenusers.Theseroutinesarerepresentedbythesimilaruserinterests,whichareidenti edbythetrajectorycorrelationalgorithm.Forexample,therecommendingsystemisabletoanswerthequestionaboutafriendwhoispassingintothecampusoftheUniversityofGrenobleduringtheweek.Therefore,thecollaborative lteringsystemveri estheuserroutines(intermsofspatio-temporalinformation)ofagroupofuserstoidentifysimilarinterestsbetweenthemFollowingthestepsofourapproachthedatasharingalgorithmcansendamessagetotheuseralertingthatafriendpassesinfrontofaspeci cnumberofthestreetXalltheweekdaysbetween10:00AMand10:30AM.Thismessagecanalsocontainaccurateinformationaboutdistance,whichisacquiredbytheHausdor distancealgorithm.The nalpartofourmiddlewareisthedata-sharingalgorithm,whichenablesthegenerationofanenrichedinformationbasedontheprocesseddata.Itreadsallthe eldsrelatedtoacorrelatedpointinordertoautomaticallycreatethemessagethatwi

35 llbesenttooneorbothusers.Figure5.13shows
llbesenttooneorbothusers.Figure5.13showsthecreationofamessagebyusingcontextinformation,whichwillbesenttotheuserBaboutapossiblepointofsocialinteractionwiththeuserA.Thedata-sharingalgorithmcanbeappliedtoseveraltypesofapplications,forexample:mobilesocialapplications,socialnetworks,SMS,andothers.Besidesthat,ourproposalallowstheinclusionofacolor-basedschemefor 5.Conclusion125 Figure5.13:ThecontextinformationofacorrelatedpointinthedatabaseoftheuserBabouttheuserA.thevisualizationofpotentialpointsofinteraction,takingintoaccounttheprobabilityofinteractionamongusers.Finally,inthenextchapter,wepresenttheevaluationofourapproach,takingintoaccountdi erentscenarios.5ConclusionVirtualcommunityplatformsprovidesolutionstosocialconnectivity,givingpeoplethecapabilitytoshareinterests,opinions,andpersonalinformationwithotherusers.Nevertheless,wearguethattheabsenceofcontext-awaremechanismsinvirtualcommunitiescouldbeoneofthemainreasonsthatsocialinteractionsarefrequentlymissed.Theusers'dailyroutines,therefore,canbecapturedbymobilesocialapplicationsandsharedinvirtualcommu-nitiesinordertoimprovethesocialconnectionsinrealcommunities.Inthischapter,weintroducedourlocation-basedapproachtoidentifysimilarinterestsbetweenusersinsocialnetworks(LIDU).Thekeyideaistoprovideamiddlewareofs

36 ervicestoacquiredailyroutinesinorderto&#
ervicestoacquiredailyroutinesinorderto ndnearpointsand,consequently,increasesocialinteractionsinrealcommunities.Wepresenteda exiblemulti-layerdatamodelformobilesocialapplicationcontextbasedonuserroutines.Wedesignedaconceptualviewtobeadapt-ableandacceptabletoasetofgenericfeaturesaswellastoassistdevelopersindesigningsolutionswiththeinherentcomplexityoftrajectorysemantics raSiPSiS  Sg Sno SaP ShII PSiP ShSSSSgSSSr IPn oS The &#xUser;&#x 55.;က&#xNumb;r 1;�A passes close to the You pass close to the address of &#xUser;&#x 55.;ကA to go raSiPSiS  Sg Sno SaP ShII PSiP SaS&#

37 x0016;S&
x0016;SSS SSSr IPn oS 126ChapterV.LIDU-Location-basedapproachtoIDentifysimilarinterestsbetweenUsersinsocialnetworks (spatio-temporaldata).Besidesthat,wediscussedhowourdatamodelcouldo ermobilesocialapplicationswithdirectsupportfortrajectories.Next,wepresentedanalgorithmtoexecutethetrajectorycorrelationprocessbasedonMinimumBoundingRectangles(MBRs)andtheHausdor distance(Haus-Dist)for ndingspatialsimilarities.Furthermore,weusedParzen-windowtechniquetoidentifysimilaritiesoftemporaldata.TovalidateourApproach,weimplementedandtestedamobilesocialapplicationfortrackingdailyroutines.Additionally,wedevelopedaplug-inonavirtualcommunityplatformtoreceivetheuserpro lesandtoexecutethetrajectorycorrelationalgorithm.Ourresultsarepresentedinthenextchapter. ChapterVIEvaluationofourapproach Contents 1Clusteringalgorithm...................1272Trajectorycorrelationalgorithm............1323Trajectorydataacquisition................1354Conclusion.........................139 Inthischapter,wepresenttheresultsobtainedbytheevaluationsofourapp

38 roachindi erentscenarios.Theseevalua
roachindi erentscenarios.Theseevaluationsweredividedinthreeparts,whichare:trajectorydataacquisition,clusteringalgorithmandtrajectorycorrelationalgorithm.Sincethemainscienti ccontributionsofthisthesisarerelatedtotheclusteringandtrajectorycorrelationalgorithms,westartpresentingthesealgorithms.Afterthat,wepresentthemobileapplicationthatwasdevelopedtoperformthetrajectorydataacquisitionprocess.Inthefollowingsectionswepresentthesepartsanddiscusstheresultsobtainedineachevaluation.1ClusteringalgorithmTodemonstratetheeciencyoftheclusteringalgorithmwehaveappliedourapproachtotwoseparateusers,basedontheirregisteredtrajectoriesinDublin,Ireland.Theoverallapproachcanbesummarizedinthreesteps.Firstofallclusteringisappliedtoindividualusertrajectoriesoveraperiodofonemonth.Auser'sdailyroutineisatrajectoryfromhometowork.Afterobtainingdistinctgroupsanaggregatedtrajectoryhastobechosen.Withthehelpofvisualizationandaggregationtechniques,abestrep-resentativetrajectoryforeachuserisobtained.Thisaggregatedtrajectory 128ChapterVI.Evaluationofourapproach obtainedfromseveralusertrajectoriesisthencomparedtootherusersbyap-plyingourtrajectorycorrelationalgorithm.Thiswillenablegroupsofuserstosharesimilarroutestoincreasegeospatialsocialinteraction.Wenowexplainthedi erentinp

39 utparameterswehaveusedinordertoverifythe
utparameterswehaveusedinordertoverifytheresults. (a)User1(=1000&minNbs=3). (b)User2(=1000&minNbs=3).Figure6.1:Reachabilityplotsshowingclusteringstructure.OPTICSclusteringalgorithmrequirestwoinputparameters:distancethreshold()andminimumneighbors(minNbs).TheauthorsofOPTICS[1]suggestthatthevalueofthesetwoparametershavetobelargeenoughtoyieldgoodresults.Westructuredourexperimentinawaythatwechoosearangeofdistancethresholdvaluesaswellasminimumneighbors.Forourscenario,wede nedthedistancethresholdbetween1000metersand15000meters)(100015000).Similarly,forminimumneighborsweselectedavalueof1upto10)(1minNbs10).Theexperimentwasrunwithacombinationofvaluesforbothparameters. 1.Clusteringalgorithm129 (a)ThreeclustersshowingdistinctroutesofUser1(over-layonmap). (b)ThreeclustersshowingdistinctroutesofUser1(withoutoverlay).Figure6.2:Clustersofuser1.Basedonthestatisticsandarangeofreachabilityplotsweobtained,wefoundthebestcombinationofvalues)(=1000&minNbs=3).Thisconditionrevealedasatisfactoryresultintermsoftheclusteringstructurefromthereachabilityplots.ThereachabilityplotsobtainedareillustratedinFigures6.1(a)and6.1(b).Theplotsshowre-orderingofobjects(trajectoriesinthedataset)onx-axiswhiley-axisdemonstratesthereachabilitydistances

40 betweentrajectories.Au-tomaticclusterext
betweentrajectories.Au-tomaticclusterextractiontechniquesfromagraphwerepresentedin[1][189].Thisdataindependentvisualizationprovidesanalystsahigh-levelunderstand-ingofclusteringstructure.Fromthesegraphsclusterscanbeidenti edbasedonGaussian-bumpsorvalleys.Asageneralruletheclusterstartsfroma 130ChapterVI.Evaluationofourapproach (a)ThreeclustersshowingdistinctroutesofUser2(over-layonmap). (b)ThreeclustersshowingdistinctroutesofUser2(withoutoverlay).Figure6.3:Clustersofuser2.steep-downareaandendsatasteep-uparea.Basedonthe rstplotinFigure6.1(a),wecanclearlyseethattherearetwodominantclustersinusertrajectories(trajectory2to13andtrajectory14to25)shownbythevalleysintheplot.Theotherclusterisagroupoftrajectories,whichdoesnotspeci callyformavalleyhowevertheyaregroupedtogetherintoonecluster.Thesecondgraph(seeFigure6.1(b))alsoshowsthreeclusterswithvaryingcardinalities(trajectory2to16,17to22and23to30).Inboththegraphs,the rsttrajectoryisconsideredasnoise(seeOPTICSalgorithm[1]).InFigures6.2(a),6.2(b),6.3(a)and6.4(b),thethreeclusters(fromboth 1.Clusteringalgorithm131 graphs)aredrawnindi erentstyles.Therepresentativeroutesforeachclusteraredrawnwithdi erentthicknessforvisualizationpurposes. (a)Bestrepresentativeaggregatedusertrajectories(user1). (b)Be

41 strepresentativeaggregatedusertrajectori
strepresentativeaggregatedusertrajectories(user2).Figure6.4:Bestrepresentativetrajectoriesofusers1and2.Theclustersshowthreedistinctroutesbothusersadoptedoveraperiodofonemonthtotravelfromhometowork.Onaverageeachusertrajectorycontainsalmost100points.Theclusteringstructurealsoformsdistinctgroupsbasedonaspeci crouteonaspeci cdayofthemonth.ForexampleinFigures6.2(a)and6.2(b),cluster2holdstrajectoriesstartingfromtrajectory14totrajectory25thatinclude11daysroutes.Forthisspeci ccasewecanacquireknowledgeaboutthepatternsrelatedwithaparticulardayofaweekoramonth.Forexample,ifweobservetheorderinwhichthetrajectorieswererecordedincaseofcluster2weobtain(1,2,3,4,7,8,9,12,13,14,15).Wecanapplyheuristicsandvisualizationtechniquessuchasheatmapsinordertogainmoreinsightsintouserbehaviors.Asapparentfromtheabovesequenceuser1alwaysfollowsasimilarorcloserouteduringatleastthreeconsecutivedaysofamonthsuchas(1,2,3),(7,8,9)and(13,14,15).Afteranalyzingtheclusteringstructurethenextstepisto ndanag-gregatedtrajectoryorabestrepresentativeofaparticularuserroute.Forthispurposewehaveappliedasimpleyetinterestingvisualizationtechnique.Whenallthreeclustersfrombothusersarevisualizedusingasinglegreyscalecolorscheme,itrevealsthemostfrequentrouteadopted.Thecolorhasto 132ChapterVI.Evalu

42 ationofourapproach beselectedinawaythati
ationofourapproach beselectedinawaythatitmustbetransparentenoughtovisualizethesechanges.ThephenomenonisillustratedinFigures6.4(a)and6.4,whereuser1anduser2bestrepresentativescanbevisualizedandextractedrespectivelyforfurtheranalysis.2TrajectorycorrelationalgorithmSincetheclusteringalgorithmrecognizesthebestrepresentativetrajectoryforeachuser,thetrajectorycorrelationalgorithmisexecuted.Forthisexample,thealgorithm rstlygeneratestheMBRsforeachbestrepresentativeusertrajectoryandidenti esthecorrelationbetweenbothMBRs.Afterthat,itcomputestheHausdor distanceofthepointsinthecorrelatedarea. Figure6.5:BestrepresentativetrajectoryofuserAincomparisontouserB.Inordertopresenttheaccuracyandeciencyofoursystemweusedacolor-basedschemetorepresentthepointsinthesameroadsegment,thenearpointsandthepointsoutofthecorrelatedarea.Figures6.5and6.7showthetrajectoryoftheusersAandBrespectivelywiththecolorsrepresentingthenearpointsbetweenthem.Thegreencolorrepresentsthesamesegmentthatisusedbybothusersfortheirdailyroutines.Thebluecolordenotesthepossiblepointsofinteraction,whichisinthecorrelatedareaamongthe Near points. Points out ofcontact area. Sharing a route segment. 2.Trajectorycorrelationalgorithm133 MBRs.Finally,theredcolorindicatesthepointsthatareoutofthecorrelateda

43 rea.Additionally,thesystemallowsthegener
rea.Additionally,thesystemallowsthegenerationofmessagesmakinguseofthecontextinformation. Figure6.6:BestrepresentativetrajectoryofuserBincomparisontouserA.Basedontheresults,weobservethatbothanalyzedusershavecommoninterestsandouralgorithmwasabletoidentifythesimilarroutinesbetweenthem.Thesesimilaritiesarepresentedaccordingtothesituationsdescribedinthelastchapter.Takingintoaccountthedi erentabstractionlevelsofourdatamodel,theseresultsillustratethecommonsegmentsofinterest(SoI)be-tweentwousers.Thisispossibleduetotheuseofenrichedinformationthatisassociatedwitheachlocationinthedatabase.Inotherwords,eachcoordi-nateisregisteredinthedatabasewithitsassociatedcontextinformation(e.g.,postaladdress,time,speedofthemovingobject,weather,etc).Therefore,thisenrichedinformationfacilitatestheidenti cationofsimilarsegmentsandcomesasanadditionalfeaturetoincreasetheaccuracyofthe nalresult.Theseresultsofourcorrelatedtrajectoryalgorithmareassociatedwithtwotrajectoriescontaininguserroutinesatthesameabstractionlevelofourmulti-layerdatamodel(seeFigure5.5).However,ouralgorithmalsoallowstoidentifysimilaruserroutinesindi erentabstractionlevels.Thatispos-sibleduetoourtopdownprocessingto ndthesimilarinterestsbetween Near Points. Points out ofcontact area. Sharing arout

44 e segment. 134ChapterVI.Evaluationofoura
e segment. 134ChapterVI.Evaluationofourapproach twotrajectories.Firstly,wecomparethehighestabstractionlevelsofbothusers,takingintoaccounttheregionaroundeachtrajectory.Sincewe ndthecorrelatedregionsofbothtrajectories,weperformthecomparisoninthenextlayerfor ndingsimilarroadsegmentsbetweenusers'trajectories,whichallowstoobtainmoredetailsaboutthetypeofsimilarity(e.g.near,sharing).Finally,wecarryoutthecomparisonatthelowestabstractionlevelinorderto ndsimilaritiesbetweenlocalplaces,suchas:bakeryX,hospitalY,andothers.Figure6.7illustratesthesamecomparison,butatadi erent(lessdetailed)abstractionlevel.TheroutineofuserBisGrenoble,sincehis/herwholetrajectoryiswithinGrenoble(Level1ofourdatamodel).Ontheotherhand,theroutineofuserAisrepresentedbyroadsegments(Level2ofourdatamodel).Basedonthat,thetrajectoryalgorithm ndsthesimilaritiesbetweentheroutinesofuserB(atthelevelofTrajectoryofInterest(ToI))incomparisontotheroutineofuserA(atthelevelofSegmentsofInterest(SoI)).AstheroutineofuserAisasubsetofthesetoftheToIrepresentedbyGrenoble,themapisshownwithagreendotoverGrenoble.Figure6.7presentsanexampleofhowamulti-layerdatamodelcanprovideinformationatdi erentabstractionlevels. Figure6.7:BestrepresentativePoI(Grenoble)ofuserBincomparisontouserAatadi erenta