萨师煊国际大数据分析与研究中心.ppt_第1页
萨师煊国际大数据分析与研究中心.ppt_第2页
萨师煊国际大数据分析与研究中心.ppt_第3页
萨师煊国际大数据分析与研究中心.ppt_第4页
萨师煊国际大数据分析与研究中心.ppt_第5页
已阅读5页,还剩35页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

WeiyiMeng孟卫一DepartmentofComputerScienceStateUniversityofNewYorkatBinghamtonJuly9 2012 Large ScaleDistributedInformationRetrievalontheWeb 萨师煊国际大数据分析与研究中心SummerResearchCampSeminar AboutSUNY Binghamton Foundedin1946afterWWII LocatedinBinghamton acityinSouthernTierofNewYorkStateAbout15 000students 3 000gradstudents IBMwasfoundedinBinghamtonOneofthe4UniversityCentersofSUNYsystem SUNYatStonyBrook SUNYatBuffalo SUNYatAlbany Formoreinformation seehttp www2 binghamton edu features premier index html WhatisInformationRetrieval Informationretrieval IR isacomputersciencedisciplineforfindingunstructureddata usuallytextdocuments thatsatisfyaninformationneedfromwithinlargecollectionsthatarestoredoncomputers Inthisseminar wearegoingtoextendthisdefinitiontoincludebothunstructuredandstructureddata WhatisDistributedInformationRetrieval DIR ItisaspecialbranchofinformationretrievalwherethedataoftheIRsystemarestoredinmultipledistributedlocations collections IntheWebenvironment DIRdealswithdatathataredistributedacrossmanywebsitesorwebservers RelatedtermsforDIR metasearchengine federatedsearch webDBintegrationsystem TheScale HowLarge ItcanbeaslargeasthenumberofdatasourcesontheWeb A2007survey Madhavanetal 2007 indicatestherewereabout50millionsearchableWebdatasourcesin2007 25millionforun orlessstructureddata webpages weibo 25millionforstructureddata webdatabases WheredoWebdatareside IcebergStructure AsmallfractionisontheSurfaceWebwithmostlystaticwebpagesthatarecrawlablebyfollowinghyperlinks Publiclyindexableportion 40 60billionpagesMostareintheDeepWebwithbothstructureddataandlessstructuredtextdocumentshiddenbehindnumeroussearchinterfaces About1trillionpages records TwoparadigmstoprovideintegratedaccesstoWebdata Crawling based GatherWebdatafromvariousWebserversand orsearchenginesandbuildasearchindexforthegathereddata SurfaceWebcrawlingDeepWebcrawlingMetasearching based DIR based Integrateexistingsearchenginesintofederatedsystems MetasearchingtextdocumentsMetasearchingstructureddatabydomain Advantagesofeachapproach Crawling based Completecontroloncrawleddata CanaddmetadataCanlinkdatafromdifferentsourcesinadvanceCancreateanarchivegraduallyCompletecontrolonretrievingtechniquesandrankingfunctionsFastresponsetime Metasearching based CapabilitiesofsearchenginescanbeleveragedNaturalclusteringofthedatabyindividualsearchenginescanbeutilizedThree levelqueryevaluationprocess SEselection SEretrieval resultmerging canleadtobettereffectivenessMorelikelytoobtainfresherresults Disadvantagesofeachapproach Crawling based DeepWebcrawlingdifficultOftenincompleteManysitesnotcrawlableLosesemantics structureofthedataCannotleveragesearchengines capabilitiesCrawlingdelayleadstolessup to dateresultsCopyrightandprivacyissues Metasearching based PerformancedependsonthequalityofusedsearchenginesMaycausesearchenginestocrashAccesscouldbeblockedbysearchenginesNodirectcontrolofthedataSlowerresponsetime Conclusions Bothtechnologies crawling basedandmetasearching based haveuniquevaluesandtheyshouldco exist Theyactuallycomplementeachother Question Isthereaneffectivewaytocombinebothtechnologiesintoasingleplatform Ourseminarwillfocusonthemetasearching DIR basedapproach Twotypesofmetasearchingsystems Becausestructuredandunstructureddatahaveverydifferentcharacteristics theyareoftenhandledseparatelywithdifferenttechnologies Metasearchingsystemsfortextdocuments metasearchenginesorDIRsystems Metasearchingsystemsforstructureddata eachforagivendomain Webdatabaseintegrationsystems Wewillfirstintroducelarge scalemetasearchenginesandthenintroducelarge scaleWebdatabaseintegrationsystems Duetolimitedtime wewillfocusonchallengesandremainingchallenges notoncurrentsolutions Large ScaleMetasearchEngines MSE useruserinterfacequerydispatcherresultmergersearchsearchsearchengine1engine2enginen texttexttextsource1source2sourcen query result AsimpleMSEarchitecture Whatisalarge scaleMSE Alarge scalemetasearchengineneedstosatisfyALLofthefollowingrequirements Itisametasearchengine Itisconnectedtoalargenumberof thousandsormore componentsearchengines Thecomponentsearchenginesarespecial purposesearchenginesCoveringaspecificdomain news sports medicine Coveringaspecificorganization RenDa IBM ACM Whythethirdrequirement ToretaintheadvantagesonfreshnessandsearchingthedeepWeb Technicalchallengeswithlarge scaleMSE ScalableandaccuratesearchengineselectionMostsearchenginesareuselessforagivenuserquery Best10results 10 000searchengines atleast9990useless UsinguselesssearchenginesisbadUnnecessarynetworktrafficWasteresourcesoflocalsearchenginesIncurhighercostatthemetasearchengineLeadtopooreffectivenessHowtoidentifythemostappropriatesearchenginesforanygivenqueryaccuratelyandinatimelymanner Howtosummarizeasearchenginecontent representative Howtocollecttherepresentative Howtousetherepresentativestoperformselection Technicalchallenges cont AutomaticsearchengineinclusionintometasearchengineAutomaticconnectiontosearchengines automaticconnectionwrappergeneration SubmitqueriesandreceiveresultpagesviaprogramAutomaticsearchresultrecords SRR extraction automaticextractionwrappergeneration AutomaticwrappermaintenanceSearchenginesmaychangetheconnectionparametersandandresultpresentationanytime Technicalchallenges cont EffectiveandefficientresultmergingAutonomouscomponentsearchengineslikelyemploydifferentmatchingtechniquesbetweenqueriesanddocuments indextechniques weightingschemes similarityfunctions link basedranking etc LocalscoresandranksaregenerallynotcomparableHowtore ranktheresultsreturnedfromdifferentsearchenginesintoasinglerankedlistsuchthathigheffectivenesscanbeachievedinaspeedymanner Large scaleMSEarchitecture SearchEnginem SearchEngineSelector QueryDispatcher ResultMerger ResultCollectorandExtractor SearchEngine1 SearchEngineRepresentatives Userquery WorldWideWeb Web SearchEngineDiscovery SEList SEIncorporation Automaticconnectionandresultextraction MetasearchEngineConstructionModule QueryProcessingModule Result SearchengineRepresentativesGeneration TwoRecentBooks Monographs W MengandC Yu AdvancedMetasearchEngineTechnology Morgan ClaypoolPublishers December2010 MetasearchenginearchitectureSearchengineselectionSearchengineincorporationResultmergingSummaryandFutureResearch TwoRecentBooks Monographs M ShokouhiandL Si FederatedSearch FoundationsandTrendsinInformationRetrieval 5 1 pp 1 102 2011 Tableofcontent IntroductionCollectionrepresentationCollectionselectionResultmergingFederatedsearchtestbedsConclusionandFutureResearchChallenges SearchEngineSelection 1 Problem Givenanyuserqueryandasetofsearchengines ordocumentcollections determinethesearchenginesthatmatchtheuserquerythebest Basicsolution Summarizethecontentofeachsearchengineinadvance Foreachuserquery compareitwiththesearchenginesummariesandcomputeamatchingscore Ranksearchenginesindescendingorderoftheirmatchingscoreswiththequeryandselectthetop rankedsearchengines SearchEngineSelection 2 Question1 Howtosummarizethecontentofeachsearchengines Advancedsolutionsarestatistics based Oneormorestatisticsforeachterminthedocumentsofasearchengine Someusedstatisticsforatermt documentfrequency df Thenumberofdocumentsinthesearchenginethatcontaint collectionfrequency cf Thenumberofsearchenginesinametasearchenginethatcontaint averagenormalizedweight anw TheavgoftheweightsoftinalldocumentscontainingtinaSE maximumnormalizedweight mnw ThemaxoftheweightsoftinalldocumentsinaSE SearchEngineSelection 3 Question2 Howtoobtainthesummariesofsearchengines Twogeneralscenarios Straightforwardcomputationifthedocumentsofthesearchengineisavailable Query basedsamplingifthedocumentsofthesearchenginearenotdirectlyavailable i e deepwebsearchengine Manypublishedsolutions butstillnotscalabletolarge scalemetasearchengines SearchEngineSelection 4 Question3 Howtoranksearchenginesforeachuserquery Sub questions Howtodefineameasureofusefulnessofasearchenginewithrespecttoaquery Howtocomputethemeasureveryquickly highlyefficiently inalarge scalemetasearchengine Alargenumberofsearchengineselectionalgorithmshavebeenproposed mostarenotveryscalable AutomaticconnectiontoanysearchenginegivenitsURLPassqueriestothesearchengineprogrammatically Receiveresultsfromthesearchengineprogrammatically AutomaticextractionofretrievedsearchresultsExtracttheURLsandsnippetsofretrievedpages ExtractthenumberofhitsExtracttheURLpatternofthenextpagebutton AutomaticconnectionandextractionmaintenanceAutomaticfailuredetection AutomaticSearchEngineIncorporation ExtractconnectionparametersfromtheHTMLformtagofeachsearchengine ApplyHTTPrequestmethod GETorPOST toperformconnection AutomaticSearchEngineConnection ComplexsearchformswithmanycontrolelementsIll formattedHTMLsearchformsMultiplesearchformsonthesamepageSearchformswithJavaScriptand orCSS cascadingStyleSheets SearchformsthathaveactionredirectionsSearchformsthatutilizesessions cookiesSearchenginesthatdonotallowmetasearching Searchformextraction Difficulties Asearchresultrecord SRR consistsofthereturnedinformationassociatedwitharetrievedWebpage URLofthepageTitleofthepageAshortsummaryofthepageOthermisc size date category Resultpagesoftencontainirrelevantinformationsuchasthatrelatedtoadvertisementandhostingorganization inadditiontoSRR AutomaticSearchResultRecords SRRs Extraction 1 WebScales WrapperGeneration anSRR anSRR ExtractcorrectSRRsfromreturnedresponsepageswhilediscardingirrelevantinformation Theproblemistoidentifytherules oftencalledwrapper thatcanextractthecorrectSRRs AutomaticSRRExtraction 2 GeneralmethodologyUtilizethetagstrings DOMtrees visualinformationononeormoreresultpagesfromthesamesearchenginetomineextractionpatterns Identifytheminimaldata richregion subtreethatlikelycontainstheSRRs Identifyseparator s thatseparatedifferentSRRs Morerecentsolutionsusemorevisualinformationonresultpages Stillcannothandlecomplexresultpageswell javascript multiplecolumns multiplesections multipleSRRformats AutomaticSRRExtraction 3 ResultMerging 1 Problem Mergereturneddocumentsfrommultiplesourcesintoasinglerankedlist DifficultiesFulldocumentsofsearchresultsarenotavailableortooexpensivetodownloadandanalyzeonthefly Localsimilarities thuslocalranks areusuallynotcomparableduetodifferentsimilarityfunctionsdifferenttermweightingschemesdifferentstatisticalvalues e g globalidfvs localidf ResultMerging 2 Alargenumberofsolutionshasbeenproposedtoperformresultmerging Someuselocalsimilaritiesassociatedwitheachresult modernsearchenginesnolongerprovidetheinformation Someuselocalranksofsearchresults Someanalyzedownloadedfulldocuments Someusethetitlesandsnippetsofthesearchresults Someconsiderthequalityoftheusedsearchengine Someconsiderwhetheraresultisretrievedfrommultiplesearchengines Someuseasamplesetofdocumentsfromeachsearchengine Informationthatcouldbeutilizedforresultmerging LocalsimilarityorlocalrankofeachresultTitleofeachresultSnippetofeachresultPublicationtimeofeachresultOrganization personwhopublishedtheresult fromURL SizeofeachresultNumberofsearchenginesthatreturnedtheresultRankingscoresofthesearchenginesthatreturnedtheresultFullcontentofeachresult orsomeoftheresults PageRankornumberofbacklinksofeachresultAsamplesetofdocumentsfromeachsearchengine ResultMerging 3 RemainingResearchChallenges 1 SearchenginesummarygenerationandmaintenanceQuery basedsamplingmethodshavenotbeenshowntobepracticallyviableforalargenumberoftrulyautonomoussearchengines Certainstatisticsusedbysomesearchengineselectionalgorithms suchasthemaximumnormalizedweight arestilltooexpensivetocollectasitmayrequiresubmittingasubstantialnumberofqueriestocoverasignificantportionofthevocabularyofasearchengine Theimportantissueofhowtoeffectivelymaintainthequalityofsummariesforsearchengineswhosecontentsmaychangeovertimehasstartedtogetattentiononlyrecentlyandmoreinvestigationintothisissueisneeded RemainingResearchChallenges 2 Automaticsearchengineconnectionwithcomplexsearchforms Moreandmoresearchenginesareemployingmoreadvancedtoolstoprogramtheirsearchforms Forexample moreandmoresearchformsnowhaveJavascripts Somesearchenginesalsoincludecookieandsessionidintheirconnectionmechanism Thesecomplexitiesmakeitsignificantlymoredifficulttoautomaticallyextractallneededconnectioninformation RemainingResearchChallenges 3 Automaticmaintenance Searchenginesusedbymetasearchenginesmaymakevariouschangesduetoupgradeorotherreasons Possiblechangesmayincludesearchformchange queryformatchange andresultdisplayformatchange Thesechangescancausethesearchenginesnotusableinthemetasearchenginesunlessnecessaryadjustmentsarem

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论