Semantic Analysis in Web Usage Mining

UniversitéLibredeBruxellesFacultédesSciencesAppliquéesDépartementInformatiqueetRéseauxSemanticAnalysisinWebUsageMiningJean-PierreNorguetPromoteur:Pr.EstebanZimányiThèseprésentéeenvuedel’obtentiondugradeacadémiquedeDocteurenSciencesAppliquéesAnnéeacadémique2005-2006AbstractWiththeemergenceoftheInternetandoftheWorldWideWeb,theWebsitehasbecomeakeycommunicationchannelinorganizations.TosatisfytheobjectivesoftheWebsiteandofitstargetaudience,adaptingtheWebsitecontenttotheusers’expectationshasbecomeamajorconcern.Inthiscontext,Webusagemining,arelativelynewresearcharea,andWebanalytics,apartofWebusageminingthathasmostemergedinthecorporateworld,offermanyWebcommunicationanalysistechniques.Thesetechniquesincludepredictionoftheuser’sbehaviourwithinthesite,comparisonbetweenexpectedandactualWebsiteusage,adjustmentoftheWebsitewithrespecttotheusers’interests,andminingandanalyzingWebusagedatatodiscoverinterestingmetricsandusagepatterns.However,WebusageminingandWebanalyticssufferfromsignificantdrawbackswhenitcomestosupportthedecision-makingprocessatthehigherlevelsintheorganization.Indeed,accordingtoorganizationstheory,thehigherlevelsintheorganizationsneedsummarizedandconceptualinformationtotakefast,high-level,andeffectivedecisions.ForWebsites,theselevelsincludetheorganizationmanagersandtheWebsitechiefeditors.Attheselevels,theresultsproducedbyWebanalyticstoolsaremostlyuseless.Indeed,mostoftheseresultstargetWebdesignersandWebdevelopers.Summaryreportslikethenumberofvisitorsandthenumberofpageviewscanbeofsomeinteresttotheorganizationmanagerbuttheseresultsarepoor.Finally,page-groupanddirectoryhitsgivetheWebsitechiefeditorconceptualresults,butthesearelimitedbyseveralproblemslikepagesynonymy(severalpagescontainthesametopic),pagepolysemy(apagecontainsseveraltopics),pagetemporality,andpagevolatility.WebusageminingresearchprojectsontheirparthavemostlyleftasideWebanalyticsanditslim-itationsandhavefocusedonotherresearchpaths.Examplesofthesepathsareusagepatternanalysis,personalization,systemimprovement,sitestructuremodification,marketingbusinessintelligence,andusagecharacterization.ApotentialcontributiontoWebanalyticscanbefoundinresearchaboutreverseclusteringanalysis,atechniquebasedonself-organizingfeaturemaps.Thistechniqueinte-gratesWebusageminingandWebcontentmininginordertoranktheWebsitepagesaccordingtoanoriginalpopularityscore.However,thealgorithmisnotscalableanddoesnotanswerthepage-polysemy,page-synonymy,page-temporality,andpage-volatilityproblems.Asaconsequence,theseapproachesfailatdeliveringsummarizedandconceptualresults.AninterestingattempttoobtainsuchresultshasbeentheInformationScentalgorithm,whichproducesalistoftermvectorsrepresentingthevisitors’needs.Thesevectorsprovideasemanticrepresentationofthevisitors’needsandcanbeeasilyinterpreted.Unfortunately,theresultssufferfromtermpolysemyandtermsynonymy,arevisit-centricratherthansite-centric,andarenotscalabletoproduce.Finally,accordingtoarecentsurvey,noWebusageminingresearchprojecthasproposedasatisfyingsolutiontoprovidesite-widesummarizedandconceptualaudiencemetrics.Inthisdissertation,wepresentoursolutiontoanswertheneedforsummarizedandconceptualaudiencemetricsinWebanalytics.WefirstdescribedseveralmethodsforminingtheWebpagesoutputbyWebservers.Thesemethodsincludecontentjournaling,scriptparsing,servermonitoring,iiiivnetworkmonitoring,andclient-sidemining.ThesetechniquescanbeusedaloneorincombinationtominetheWebpagesoutputbyanyWebsite.Then,theoccurrencesoftaxonomytermsinthesepagescanbeaggregatedtoprovideconcept-basedaudiencemetrics.Toevaluatetheresults,weimplementaprototypeandrunanumberoftestcaseswithrealWebsites.AccordingtothefirstexperimentswithourprototypeandSQLServerOLAPAnalysisService,concept-basedmetricsproveextremelysummarizedandmuchmoreintuitivethanpage-basedmet-rics.Asaconsequence,concept-basedmetricscanbeexploitedathigherlevelsintheorganization.Forexample,organizationmanagerscanredefinetheorganizationstrategyaccordingtothevisitors’interests.Concept-basedmetricsalsogiveanintuitiveviewofthemessagesdeliveredthroughtheWebsiteandallowtoadapttheWebsitecommunicationtotheorganizationobjectives.TheWebsitechiefeditoronhispartcaninterpretthemetricstoredefinethepublishingordersandredefinethesub-editors’writingtasks.Asdecisionsathigherlevelsintheorganizationshouldbemoreeffective,concept-basedmetricsshouldsignificantlycontributetoWebusageminingandWebanalytics.AcknowledgementsThewritingofthisdissertationhasbeendirectedbyprofessorEstebanZimányi.Hissupport,hiscomments,andhissenseofbalancemadepossibletherealizationofthiswork.Iamalsogratefulforthepatiencehehasdeployedtoeducateme,notonlytoscientificresearchbutalsotocriticism,andtoanotherperceptionofreality.Ialsothankthemembersofmycommitteefortheirattentionandcomments.ThecommitteewascomposedofPr.HuguesBersiniaspresident,Pr.RoelWuytsassecretary,Pr.EstebanZimányiasdissertationdirector,Pr.GianlucaBontempiaspresidentoftheaccompanimentcommittee,andPr.Marie-ChristineRoussetfromUniversityofGrenobleasexternalexpert.Inparticular,Iwanttoacknowledgethebrillanceandmasteringo