Presentation on theme: "BANKS Browsing and Keyword Searching in Relational Databases B"— Presentation transcript
BANKS:BrowsingandKeywordSearchinginRelationalDatabasesB.AdityaGauravBhalotiaSoumenChakrabartiArvindHulgeriCharutaNakheParagS.SudarshanComputerScienceandEngg.Dept.,I.I.T.Bombaybaditya,soumen,aru,parag,sudarsha@cse.iitb.ac.inbhalotia@eecs.berkeley.educharuta@pspl.co.inAbstractTheBANKSsystemenableskeyword-basedsearchondatabases,togetherwithdataandschemabrowsing.BANKSenablesusersto
PartlysupportedbyanIBMFacultyFellowshipgrantand
Mining Surprising Patterns UsingTemporal Description Length
ChakrabartiSD98...Paper Tuple
PaperIdPaperName
ByronDChakrabartiSD9
8
Writes Tuple
SunitaSChakrabartiSD98Writes Tuple
...Sunita SarawagiSunitaSAuthor Tuple
A
uthorIdAuthorName
.
..
Byron DomByronDAuthor Tuple
Figure1:AFragmentoftheDBLPDatabaseBANKS,includingrelevancescorecomputation,isde-scribedinmoredetailinSection2.BANKSprovidesarichinterfacetobrowsedata,withautomaticgenerationofhyperlinks.Thebrows-ingcomponentofBANKSisdescribedinmorede-tailinSection3.TheBANKSsystemisdevelopedinJavausingservletsandJDBC,andcanberunonanydatabase,withoutanyprogramming.Wearealsode-velopingaversionofBANKSthathandlesXMLdata.ThegreatestvalueofBANKSliesinalmostzero-eortWebpublishingofrelationaldatawhichwouldotherwiseremaininvisibletotheWeb[2].Forexam-ple,BANKSmaybeusedtopublishorganizationaldata,bibliographicdata,andelectroniccatalogs.AdemooftheBANKSsystemisaccessibleovertheWebattheURL:http://www.cse.iitb.ac.in/banks/2KeywordsearchConsiderafragmentofabibliographicdatabaseshowninFigure1.Thisdatabasecontainspapertitles,theirauthorsandcitationsextractedfromtheDBLPreposi-tory.Aswecansee,duetonormalization,informationaboutasinglepaperisdistributedacrossseventuplesrelatedthroughforeignkeyreferences.AuserlookingforthispapermayusequerieslikesunitatemporalŽorsoumensunitaŽ.TheBANKSsystemmodelsthedatabaseasadi-rectedgraph,witheachtupleinthedatabasecorre-spondingtoanodeinthegraph.Eachforeign-key…primary-keylinkismodeledasadirectededgebetweenthecorrespondingtuples.(Thiscanbeeasilyextendedtoothertypeofconnections.)AkeywordqueryinBANKSconsistsof1searchterms,...,t.The“rststepistolocatethesetofnodesmatchingsearchterms;anodematchesasearchtermifitcontainsthesearchtermaspartofanat-tributevalueormetadata(suchascolumn,tableorviewnames).Letdenotethesetofnodesmatch-
Figure2:ResultofquerysoumensunitaŽ;anodemaymatchmorethanonesearchterm,sothesmayoverlap.Intuitively,ananswertoaqueryisasubgraphconnectingsomesetofnodesthatcoverŽthekeywords,i.e.,eachkeywordmustmatchoneofthenodesin.Justbylookingatasubgraphitmaynotbeapparentwhatinformationitconveys.WewishtoalsoidentifyacentralŽnodeinthesubgraph,thatconnectsallthekeywordnodes,andstronglyre”ectstherelationshipamongstthem.Ananswertoaqueryisthereforemodeledasarooteddirectedtreecontainingatleastonenodefromeach;edgesaredirectedawayfromtheroot.Themotivationfordirectionalityisoutlinedlaterinthissection.NotethatthetreemayalsocontainnodesnotinanyandisthereforeaSteinertree.Fig-ure2showsasampleresultofaquerycontainingthekeywordssoumenandsunitaexecutedonthebiblio-graphicdatabase.Indentationisusedtodepictthetreestructure,andnodescontainingkeywordsaredis-tinguishedbytheircolor.2.1AnswerRelevanceIngeneral,theimportanceofalinkdependsuponthetypeofthelink,i.e.,whatrelationsitconnectsandonitssemantics;forexample,inabibliographicdatabase,thelinkbetweenthePapertableandtheWritestableisseenasastrongerlinkthanthelinkbetweenthePapertableandtheCitestable.ThelinkbetweenPaperandCitestablescancorrespondinglybegivenahigherweight.Theweightofatreeisproportionaltothetotalofitsedgeweights,andtherelevanceofatreeisinverselyrelatedtoitsweight.TheexampleinFigure1illustratesthatsomelinkspointtowardtherootofthetree,insteadofawayfromtherootasrequiredbyourmodel.Forinstance,theWritesrelationhasforeignkeystothePaperandAu-thorrelations,whereaswerequirepathsfromPaper
Author,traversingaforeignkeyedgeintheop-positedirection.However,wecannotsimplyregardtheedgesasundirected.IgnoringdirectionalitywouldcauseproblemsbecauseofhubsŽthatareconnectedtoalargenumbersofnodes.Forexample,inauni-versitydatabaseadepartmentwithalargenumberoffacultyandstudentswouldactasahub.Asaresult,manynodeswouldbewithinashortdistanceofmanyothernodes,reducingtheeectivenessoftree-weightbasedscoringmechanism.Tosolvetheproblem,wecreateforeachedge(u,vbackwardedgev,u);intheexamplefromFigure1,thebackwardedgesensurethatthereisadirectedtreerootedatthepaper,withapathtoeachleaf.Wesettheweightof(v,u)totheweightof(u,v)multipliedbyafunctionofthenumberoflinkstofromthenodesofthesametypeas.Experimentswithdierentfunctionsindicatedthatthefunctionlog(1+),whereisthenumberofinlinks,providedgoodresults[3].(Iftherewasalreadyanedgefrom,wesettheedgeweighttotheloweroftheoriginaledgeweightandtheweightcomputedabove.)BANKSincorporatesanotherinterestingfeature,namelynodeweights,inspiredbyprestigerankingssuchasPageRankinGoogle[4].Withthisfeature,nodesthathavemultiplepointerstothemgetahighernodeweight(highernodeweightcorrespondstohigherprestige).E.g.,inabibliographydatabasecontainingcitationinformation,iftheusergivesaqueryQueryOptimizationourtechniquewouldgivehigherprestigetothepaperswithmorecitations.Asanotherexam-ple,inaTPCDdatabasestoringinformationaboutparts,suppliers,customersandorders,theordersin-formationcontainsreferencestoparts,suppliersandcustomers.Asaresult,ifaquerymatchestwoparts(orsuppliers,orcustomers)theonewithmoreorderswouldgetahigherprestige.Inthecurrentimplementationwesetthenodeweighttoafunctionofthein-degreeofthenode.Weexperimentedwithdierentfunctions,andgotgoodresultswiththefunctionlog(1+),whereisthein-degree.Ouruseoflogarithmsinedgeandnodeweightsissimilartotermweightingschemesininfor-mationretrieval.Nodeweightsandtreeweightsarecombinedtogetanoverallrelevancescore.Weexperimentedwithad-ditiveandmultiplicativecombinations,andfoundthatbothworkedwellwhentherelativeweightsforthetwoscoreswereappropriatelychosen.Detailsofthesearchalgorithmandtherelevancecomputation,alongwithapreliminaryperformancestudycanbefoundin[3].Althoughafewothersystemsimplementkeywordsearchondatabases(e.g.,[5,1,6])BANKSdiersfromallpriorworkinseveralways:notably,inthetech-niquesforedgeweightcomputationandprestigebasedranking,andtheuseofanin-memorygraphstructureforveryecientsearchwhilekeepingthebulkofthedatadiskresident.TheconnectionsofBANKStore-latedworkaredescribedinmoredetailin[3].2.2ExtensionsTheBANKSsystemsupportsiterativere“nementofIfmultiplenodesmatchakeyword,theusercanselectoneormorenodesasbeingrelevantandignoreothers;asanexample,twoauthorsintheDBLPdatabasematchthekeywordsudarshanŽ,andtheusercanchooseoneofthemandexecuteaquerymatchingitwithotherkeywords.Userscanrequestmoreanswerssimilartooneofthedisplayedanswers;similaritycanbede“nedonthebasisoftheanswertreestructure.Otherre“nementstotunenodeandedgeweightsarealsounderdevelopment.Insteadofdisplayingtreesconsistingofexplicittu-ples,systemdeveloperscanspecifyanswerformatsbasedonthetypeoftherootoftheanswertrees.Forinstance,onecanspecifyauthor,conference/journalandyearbedisplayedwhenevertherootnodeisfrompaperrelation.Wearecurrentlyworkingonimple-mentinganswerformatting,andonsupportingnega-tionanddisjunctioninqueries.3BrowsingTheBANKSsystemprovidesarichinterfacetobrowsedatastoredinarelationaldatabase.Thesystemau-tomaticallygeneratesbrowsableviewsofdatabasere-lationsandqueryresults;nocontentprogrammingoruserinterventionisrequired.Everydisplayedforeignkeyattributevaluebecomesahyperlinktothereferencedtuple.Inaddition,pri-marykeycolumnscanbebrowsedbackwards,to“ndreferencingtuples,organizedbyreferencingrelations(userscanselectaspeci“creferencingrelation).Eachtabledisplayedcomeswithavarietyoftoolsforinteractingwithdata.Columnscanbeprojectedaway(dropped),andselectionscanbeimposedonanycolumn.Forforeignkeycolumns,clickingonjoinŽresultsinthereferencedtablebeingjoinedin,anditscolumnsalsodisplayed.Thiseliminatestheneedforexplicitlywritingjoinqueriesforthenormalcaseofforeignkeyjoin.Thejoinfeaturecanalsobeusedintheotherdirection,fromaprimarykeytoareferencingforeignkey.Resultscanbegroupedbyacolumn;thisresultsinonlythedistinctvaluesforthatcolumnbeingdisplayed.Theusercanclickonanyofthevaluestoseethetuplesassociatedwiththatvalue.Tuplesinthedisplayedtablecanbesortedbyaspeci“edcolumn.
Figure3:BrowsingExamples:(a)Samplebrowsingsession(b)PiechartControlsfortheseoperationscanbeaccessedbyclickingonthecolumnnamesinthetableheader.Inaddition,displayeddataispaginated,andschemabrowsingissupported.Figure3(a)showstheresultofbrowsingthethe-sisdatabasestartingwiththestudentrelation,usingapop-upmenuontherollnumberattributetoeectajoinwiththethesisrelationanddroppingseveralcolumns.Thejoinismadepossiblesincethethesislationhasaforeign-keyattributereferencingthedentrelation.Asamplepop-upmenuisshownforthefemailattributewhichreferencesthefacultytable.Hyperlinksinthedisplayeddataareautomaticallygeneratedbythesystem.EachhyperlinkcorrespondstoanSQLquerythatisexecutedwhenauserclicksonthelinks.Thus,allthepagesinthesystemaregen-eratedonthe”ybyexecutingcorrespondingqueriesagainsttheunderlyingrelationaldatabase.Nopre-computationisrequired.BANKStemplatesprovideseveralprede“nedwaysofdisplayingdata.Templateinstancesarecustomized,storedinthedatabase,andgivenahyperlinkname,whichisusedtoaccessthetemplate.TheBANKSsystemcurrentlyprovidesfourtypesoftemplates:Cross-tabs(similartoOLAPcross-tabs),withdrill-downfacilities.Thegroup-bytemplateprovidesahierarchicalviewofdata,byspecifyingasequenceofgroupingattributes.Folder-treeviews,whichprovideanotherhierar-chicalviewofdata.Thegraphicalinterfacetemplatepermitsinforma-tiontobedisplayedinbarchartlinechartpiechartformat.HyperlinksareprovidedonthegraphicaldataviaHTMLimagemaps,toallowdrilldownonthedata.Figure3(b)showsanex-amplepiechartgeneratedbyBANKS.Anotherinterestingfeatureoftemplatesisthattheycanbecomposedtogetherinahyperlinked,visualmanner.SeveralexampletemplatesareavailableontheBANKSwebsite.4ConclusionsTosummarize,wehavedevelopedanintegratedsys-temforkeywordsearchingandbrowsingofdatabases.Thesystemhasmanyusefulfeatureswhichallowca-sualuserstoaccessdatabaseinformationinanintu-itivemanner.BANKSenablesalmosteortlessWebpublishingofrelationalandXMLdatathatwouldoth-erwiseremain(atleastpartially)invisibletotheWeb.WehavealsodevelopedaprototypeversionofBANKSthatworksonXMLdata,supportingkey-wordsearchingandscalablebrowsingoflargeXMLdatasets.WeareworkingonintegratingXMLsearch/browsingwiththerestoftheBANKSsystem.References[1]SanjayAgrawal,SurajitChaudhuri,andGautamDas.DBXplorer:Asystemforkeyword-basedsearchoverrelationaldatabases.InProcs.ICDEFeb.2002.[2]PeterBailey,NickCraswell,andDavidHawking.DarkmatterontheWeb.InPosterProceedings,9thWorld-WideWebConference,2000.[3]GauravBhalotia,ArvindHulgeri,CharutaNakhe,SoumenChakrabarti,andS.Sudarshan.Key-wordsearchingandbrowsingindatabasesusingBANKS.InProcs.ICDE,Feb.2002.[4]SergeyBrinandLawrencePage.Theanatomyofalarge-scalehypertextualWebsearchengine.Com-puterNetworksandISDNSystems,30(1…7),1998.[5]ShaulDar,GadiEntin,ShaiGeva,andEranPal-mon.DTLsDataSpot:Databaseexplorationus-ingplainlanguage.InProcs.VLDB,1998.[6]VagelisHristidisandYannisPapakonstanti-nou.DISCOVER:Keywordsearchinrelationaldatabases.InProcs.VLDB,Aug.2002.
