The Role of Semantic Relevance in Dynamic User Com

TheRoleofSemanticRelevanceinDynamicUserCommunityManagementandtheFormulationofRecommendationsNickPapadopoulos1,2,DimitrisPlexousakis1,21DepartureofComputerScience,UniversityofCrete,P.O.Box2208,GR-71409,Heraklion,GREECE{nickpap,dp}@csd.uch.gr2InstituteofComputerScience,FoundationforResearchandTechnology-Hellas.P.O.Box1385,GR-71110,Heraklion,GREECE{npap,dp}@ics.forth.grAbstract.Intherecentyears,anincreasinginterestinrecommendationsystemshasemergedbothfromtheresearchandtheapplicationpointofviewandinbothacademicandcommercialdomains.Themajorityofcomparisontechniquesusedforformulatingrecommendationsarebasedonset-operationsoveruser-suppliedtermsorinternalproductcomputationsonvectorsencodinguserpref-erences.Inbothcaseshowever,theidentical-nessoftermsisexaminedratherthantheiractualsemanticrelevance.Thispaperproposesarecommendational-gorithmthatisbasedonthemaintenanceofuserprofilesandtheirdynamicad-justmentaccordingtotheusers’behavior.Moreover,thisalgorithmreliesonthedynamicmanagementofcommunities,whichcontainsimilarandrelevantusers.Thesecommunitiesarecreatedaccordingtoaproposedalgorithmforclassificationofrelevantusersintothecommunities.Thealgorithmisimple-mentedontopofacommunitymanagementmechanismthathasbeendevel-oped.Thecomparisonmechanismusedinthecontextofthisworkisbasedonsemanticrelevancebetweenterms,whichisevaluatedwiththeuseofaglossaryofterms.1IntroductionIntherecentyearsanincreasinginterestinrecommendationsystemshasemergedbothfromtheresearchandtheapplicationpointofview,andinbothacademicandcommercial/economicdomain.Manyonline“e-shops”haveadoptedrecommendationtechniquestorecommendnewitemstotheircustomers,basedonaloggedhistoryofpreviouspurchasesortransactions.Themajorityofexistingrecommendationsystemsdoesnotadequatelyaddresstheinformationfilteringneedsoftheirusers.Themainreasonfortheexpansionofsuchsystemsisthenecessityofamechanismabletofilterinformation,thatisavailablethroughtheWeb,anddynamicallyadjustittotheuser’sneedsandinterests[1].Informationfilteringanditssubsequenttailoringtotheuser’sinterestsconstitutetheultimategoalsofrecommendationsystems.Modernrecommendationsystemsmainlybasetheirfunctionalityononeoftwoapproaches,inordertosug-gest/recommendadocumentorproducttoauser.Thesetwoapproachesarecontent-basedfilteringandcollaborative-filtering[2,3,6,7]respectively.Theadoptionoftheformerorthelatterdependsonthetypeofinformationthatasystemisaimingtopro-videtoitsusers.1Inthecontent-basedapproach,thesystemfiltersinformation(typicallyadocument)andaimstoproviderecommendationsbasedonthecontentsofthisdocument.Hence,giventheirpreferences,recommendationsfordocumentsthatarerelevanttotheirinterestsareforwardedtotheusers[2,3].Suchanapproachhasbeenmainlyadoptedbyinformationretrieval[7,8]andmachinelearningsystems[9,10,11].Forinstance,inthecaseoftextdocuments,recommendationsareprovidedbasedonthedegreeofmatchingbetweenthecontentofadocumentandauser’sprofile.Theuser’sprofileisbuiltandmaintainedaccordingtoananalysisthatisappliedtothecontentsofthedocumentsthattheuserhaspreviouslyrated[11,12].Theuser’sratingsincombina-tionwithamechanismforobtainingthefeaturetermsofadocument[2,3,4,5]aretheonlyrequirementofthesesystems.Ontheotherhand,systemsadoptingthecollaborative-filteringapproach,aimtoidentifyusersthathaverelevantinterestsandpreferenceswithaparticularuser.Thereafter,thedocumentsthattheseuserspreferarerecommendedtothatparticularuser.Theideabehindthisapproachisthatitmaybeofbenefittoone’ssearchforinformationtoconsultthebehaviorofotheruserswhosharethesameorrelevantinterestsandwhoseopinioncanbetrusted[4].Suchsystemstakeadvantageofthedocumentsthatotherusershavealready“discovered”andrated.Inorderforthesesystemstobeableto“detect”therelevancebetweenusers,theremustbeacompari-sonmechanism,whichinthiscasetoo,istheusageofuserprofileinformation[11,12].Therequirementimposedbysuchsystemsisthateachnewuserhastoratesomedocuments,sothataprofilecanbebuiltforher[2,3,4,5].Ifthesetwoapproachesareappliedseparately,theypresentcrucialdisadvantagesandsufferfromspecificproblems.Inorderfortheseproblemstobedealtwith,hybridapproachesneedtobedevised.Motivatedbytheseissues,thispaperpresentssuchahybridapproachinwhichthesemantic(andnotsolelylexical)relevanceoftermsindynamicallymaintaineduserprofilesisexploitedfortheformulationofrecommenda-tions.Therestofthepaperisorganizedasfollows.Section2presentsacomparisonbe-tweencontent-basedfilteringandcollaborativefilteringaswellasthebenefitsofhy-bridapproaches.Section3describestheproposedalgorithmforformulatingrecom-mendations.Section4presentspreliminaryexperimentalresults.Section5summarizesourcontributionanddrawsdirectionsforfurtherresearch.1Amoredetailedcomparisonofthetwoapproachescanbefoundinsection2.2ComparingContent-BasedAndCollaborativeFilteringArepresentationofthedataavailableinarecommendationsystemisshowninfigure1.Dataarerepresentedbymeansofamatrixwhoserowscorrespondtousersandcolumnstoitems.Eachitemisassociatedwithafeaturevector.Theelementsofthismatrixarebinary,indicatingwhetherhowauserratedanitem.Fig.1.Representationofdata