ICDE论文集总结.docx
- 文档编号:11046013
- 上传时间:2023-02-24
- 格式:DOCX
- 页数:29
- 大小:34.36KB
ICDE论文集总结.docx
《ICDE论文集总结.docx》由会员分享,可在线阅读,更多相关《ICDE论文集总结.docx(29页珍藏版)》请在冰豆网上搜索。
ICDE论文集总结
2014-ICDE论文集
ICDERESEARCHSESSIONS
ResearchPapersSession1Clustering
●IncrementalClusterEvolutionTrackingfromHighlyDynamicNetworkData(shxy)
oPeiLee*(UBC)
oLaksV.S.Lakshmanan(UBC)
oEvangelosMilios(DalhousieUniversity)
摘要:
Dynamicnetworksarecommonlyfoundinthecurrentwebage.Inscenarioslikesocialnetworksandsocialmedia,dynamicnetworksarenoisy,areoflarge-scaleandevolvequickly.Inthispaper,wefocusontheclusterevolutiontrackingproblemonhighlydynamicnetworks,withclearapplicationtoeventevolutiontracking.Thereareseveralpreviousworksondatastreamclusteringusinganode-by-nodeapproachformaintainingclusters.However,handlingofbulkupdates,i.e.,asubgraphatatime,iscriticalforachievingacceptableperformanceoververylargehighlydynamicnetworks.Weproposeasubgraph-bysubgraphincrementaltrackingframeworkforclusterevolutioninthispaper.Toeffectivelyillustratethetechniquesinourframework,wetaketheeventevolutiontrackingtaskinsocialstreamsasanapplication,whereasocialstreamandaneventaremodeledasadynamicpostnetworkandadynamicclusterrespectively.Bymonitoringthroughafadingtimewindow,weintroduceaskeletalgraphtosummarizetheinformationinthedynamicnetwork,andformalizeclusterevolutionpatternsusingagroupofprimitiveevolutionoperationsandtheiralgebra.Twoincrementalcomputationalgorithmsaredevelopedtomaintainclustersandtrackevolutionpatternsastimerollsonandthenetworkevolves.OurdetailedexperimentalevaluationonlargeTwitterdatasetsdemonstratesthatourframeworkcaneffectivelytrackthecompletesetofclusterevolutionpatternsinthewholelifecyclefromhighlydynamicnetworksonthefly.
大意:
动态网络(如社交网络)在网络时代非常常见。
本文主要解决的是在动态网络上cluster演化的跟踪问题(事件的演化跟踪)。
在本文之前的跟踪方法(节点的更新)涉及到很多的更新操作。
本文提出了一个基于子图的增量追踪框架。
通过对社交流(skeletalgraph)和事件的单独建模来解决此问题。
●FindingCommonGroundamongExperts’OpinionsonDataClustering:
withApplicationsinMalwareAnalysis
oGuanhuaYan*(LosAlamosNationalLab)
摘要:
Dataclusteringisabasictechniqueforknowledgediscoveryanddatamining.Asthevolumeofdatagrowssignificantly,dataclusteringbecomescomputationallyprohibitiveandresourcedemanding,andsometimesitisnecessarytooutsourcethesetaskstothirdpartyexpertswhospecializeindataclustering.Thegoalofthisworkistodeveloptechniquesthatfindcommongroundamongexperts'opinionsondataclustering,whichmaybebiasedduetothefeaturesoralgorithmsusedinclustering.Ourworkdiffersfromthelargebodyofexistingapproachestoconsensusclustering,aswedonotrequirealldataobjectsbegroupedintoclusters.Rather,ourworkismotivatedbyreal-worldapplicationsthatdemandhighconfidenceinhowdataobjects-iftheyareselected-aregroupedtogether.WeformulatetheproblemrigorouslyandshowthatitisNP-complete.Wefurtherdevelopalightweighttechniquebasedonfindingamaximumindependentsetina3-uniformhypergraphtoselectdataobjectsthatdonotformconflictsamongexperts'opinions.Weapplyourproposedmethodtoareal-worldmalwaredatasetwithhundredsofthousandsofinstancestofindmalwareclustersbasedonhowmultiplemajorAV(Anti-Virus)softwareclassifythesesamples.Ourworkoffersanewdirectionforconsensusclusteringbystrikingabalancebetweentheclusteringqualityandtheamountofdataobjectschosentobeclustered.
大意:
聚类是知识发现和数据挖掘的一项基本技术,但是随着数据量的增多,聚类变得(computationallyprohibitive),并且有时需要外包给第三方的数据聚类专家。
本文的目标是发现专家观点的共同点。
本文的的目标是形成高可信度的聚类结果(不与专家的观点相冲突)。
●TowardsEffectiveandEfficientMiningofArbitraryShapedClusters
oHaoHuang*(NUS)
oYunjunGao(ZhejiangUniversity)
oKevinChiew(TanTaoUniversity)
oLeiChen(HongKongUniversityofScienceandTechnology)
oQinmingHe(ZhejiangUniversity)
摘要:
Miningarbitraryshapedclustersinlargedatasetsisanopenchallengeindatamining.Variousapproachestothisproblemhavebeenproposedwithhightimecomplexity.Tosavecomputationalcost,somealgorithmstrytoshrinkadatasetsizetoasmalleramountofrepresentativedataexamples.However,theiruser-definedshrinkingratiosmaysignificantlyaffecttheclusteringperformance.Inthispaper,wepresentCLASPaneffectiveandefficientalgorithmforminingarbitraryshapedclusters.Itautomaticallyshrinksthesizeofadatasetwhileeffectivelypreservingtheshapeinformationofclustersinthedatasetwithrepresentativedataexamples.Then,itadjuststhepositionsoftheserepresentativedataexamplestoenhancetheirintrinsicrelationshipandmaketheclusterstructuresmoreclearanddistinctforclustering.Finally,itperformsagglomerativeclusteringtoidentifytheclusterstructureswiththehelpofamutualk-nearestneighbors-basedsimilaritymetriccalledPk.Extensiveexperimentsonbothsyntheticandrealdatasetsareconducted,andtheresultsverifytheeffectivenessandefficiencyofourapproach.
大意:
在大数据集下挖掘任意形状的聚类有很大难度。
许多算法的时间复杂度过高,有些算法为了减少时间复杂度通过用户自定义的比例缩小数据集,从而获得代表性的数据(影响最终结果)。
本文提出CLASP算法,自动缩小数据集大小,并且保留形状信息。
并且调整数据的位置提高固有的联系。
ResearchPapersSession2DistributedProcessing
●R-Store:
AScalableDistributedSystemforSupportingReal-timeAnalytics
oFengLi*(NationalUniversityofSingapore)
oTamerOzsu(UniversityofWaterloo)
oGangChen(ZhejiangUniversity)
oBengChinOoi(NationalUniversityofSingapore)
摘要:
ItiswidelyrecognizedthatOLTPandOLAPquerieshavedifferentdataaccesspatterns,processingneedsandrequirements.Hence,theOLTPqueriesandOLAPqueriesaretypicallyhandledbytwodifferentsystems,andthedataareperiodicallyextractedfromtheOLTPsystem,transformedandloadedintotheOLAPsystemfordataanalysis.Withtheawarenessoftheabilityofbigdatainprovidingenterprisesusefulinsightsfromvastamountsofdata,effectiveandtimelydecisionsderivedfromreal-timeanalyticsareimportant.Itisthereforedesirabletoprovidereal-timeOLAPqueryingsupport,whereOLAPqueriesreadthelatestdatawhileOLTPqueriescreatethenewversions.Inthispaper,weproposeR-Store,ascalabledistributedsystemforsupportingreal-timeOLAPbyextendingtheMapReduceframework.Weextendanopensourcedistributedkey/valuesystem,HBase,astheunderlyingstoragesystemthatstoresdatacubeandreal-timedata.Whenreal-timedataareupdated,theyarestreamedtoastreamingMapReduce,namelyHstreaming,forupdatingthecubeonincrementalbasis.Basedonthemetadatastoredinthestoragesystem,eitherthedatacubeorOLTPdatabaseorbothareusedbytheMapReducejobsforOLAPqueries.Weproposetechniquestoefficientlyscanthereal-timedatainthestoragesystem,anddesignanadaptivealgorithmtoprocessthereal-timequerybasedonourproposedcostmodel.Themainobjectivesaretoensurethefreshnessofanswersandlowprocessinglatency.TheexperimentsconductedontheTPC-Hdatasetdemonstratetheeffectivenessandefficiencyofourapproach.
大意:
●Blazes:
CoordinationAnalysisforDistributedPrograms
oPeterAlvaro*(UCBerkeley)
oNeilConway(UCBerkeley)
oJoeHellerstein(UCBerkeley)
oDavidMaier(PortlandStateUniversity)
●QueryOptimizationofDistributedGraphPatternMatching
oJiewenHuang*(YaleUniversity)
oDanAbadi(YaleUniversit)
ResearchPapersSession3DataMiningI:
OutliersandTimeSeries
●ScalableDistance-BasedOutlierDetectionoverHigh-VolumeDataStreams
oLeiCao*(WPI)
oElkeRudensteiner(WPI)
oQingyangWang(WPI)
●DiscriminativeFeaturesforIdentifyingandInterpretingOutliers
oXuan-HongDang*(AarhusUniversity)
oIraAssent(AarhusUniversity)
oRaymondT.Ng.(UniversityofBritishColumbia,Canada)
oArthurZimek(LMUMunich)
oErichSchubert(Ludwig-MaximiliansUniversitätMünchen)
大意:
Weconsidertheproblemofoutlierdetectionandinterpretation.Whilemostexistingstudiesfocusonthefirstproblem,wesimultaneouslyaddresstheequallyimportantchallengeofoutlierinterpretation.Weproposeanalgorithmthatuncoversoutliersinsubspacesofreduceddimensionalityinwhichtheyarewelldiscriminatedfromregularobjectswhileatthesametimeretainingthenaturallocalstructureoftheoriginaldatatoensurethequalityofoutlierexplanation.Ouralgorithmtakesamathematicallyappealingapproachfromthespectralgraphembeddingtheoryandweshowthatitachievesthegloballyoptimalsolutionfortheobjectiveofsubspacelearning.Byusinganumberofreal-worlddatasets,wedemonstrateitsappealingperformancenotonlyw.r.t.theoutlierdetectionratebutalsow.r.t.thediscriminativehuman-interpretablefeatures.Thisisthefirstapproachtoexploitdiscriminativefeaturesforbothoutlierdetectionandinterpretation,leadingtobetterunderstandingofhowandwhythehiddenoutliersareexceptional.
大意:
Discriminative(有区别力的)本文提出的算法不仅检测离群点并且指出哪些变量最能区别这些离群点。
spectralgraph(用SVD奇异值分解进行降维)
●Memory-efficientCentroidDecompositionforLongTimeSeries
oMouradKhayati*(UniversityofZurich)
oMichaelBoehlen(UniversityofZuerich)
oJohannGamper(FreeUniversityofBozen-Bolzano)
ResearchPapersSession4ParetoOptimization
●IncrementalDiscoveryofProminentSituationalFacts(send)
oAfrozaSultana*(Univ.ofTexasatArlington)
oNaeemulHassan(Univ.ofTexasatArlington)
oChengkaiLi(Univ.ofTexasatArlington)
oJunYang(DukeUniversity)
oCongYu(GoogleResearch)
大意:
Westudythenovelproblemoffindingnew,prominentsituationalfacts,whichareemergingstatementsaboutobjectsthatstandoutwithincertaincontexts.Manysuchfactsarenewsworthy-e.g.,anathlete'soutstandingperformanceinagame,oraviralvideo'simpressivepopularity.Effectiveandefficientidentificationofthesefactsassistsjournalistsinreporting,oneofthemaingoalsofcomputationaljourna
- 配套讲稿:
如PPT文件的首页显示word图标,表示该PPT已包含配套word讲稿。双击word图标可打开word文档。
- 特殊限制:
部分文档作品中含有的国旗、国徽等图片,仅作为作品整体效果示例展示,禁止商用。设计者仅对作品中独创性部分享有著作权。
- 关 键 词:
- ICDE 论文集 总结
![提示](https://static.bdocx.com/images/bang_tan.gif)