01 宋国杰 - 北京大学信息科学技术学院副教授 - Network Representation Learning.pdfVIP免费

NetworkRepresentationLearning宋国杰北京大学信息科学技术学院2Outline➢Background➢OurRecentWorks➢AboutFutureInterconnectedWorldWORLDWIDEWEBELECTRICALNETWORKSSOCIALNETWORKSTRANSPORTATIONNETWORKSINFORMATIONNETWORKSPROTEININTERACTIONS3NodeimportanceCommunitydetectionNetworkdistanceLinkpredictionNodeclassificationNetworkevolution…4NetworkAnalyticsDescriptiveandPredictiveNETWORKANALYTICSTECHNOLOGYHAVEVITALIMPACTONMANYREALAPPLICATIONS.45ComputingPower:exponentiallygrowRequired:doubleexponentialRealBigDataproblemexistsinNetworkedData.ButItisveryHard56ButWhy?G=(V,E)Iterative&CombinatorialComplexityCouplingParallelizabilityDependencyamongnodesInapplicabilityofMLmethodsLinks67OptionalSolutionG=(V,E)G=(V)VectorSpaceEasytoparallelCanapplyclassicalMLmethodsgenerateembed7NetworkRepresentationLearning➢WhatistherelationshipbetweenNetworkembeddingandGraphNeuralNetwork?8Goal:Needtodefine!NetworkEmbeddingunsupervised9Representativemodel➢TraditionallineardimensionreductionmodelsPCA,SVD,MDSetc,Timecomplexityisatleastquadratic➢ManifoldlearningLLE,ISOMAP10LLE（Locallylinearembedding）ISOMAPWord2vecbasedNEModels11atureLearningforNetworksJureLeskovecStanfordUniversityjure@cs.stanford.edul--s-es-sus3s2s1s4s8s9s6s7s5BFSDFSFigure1:BFSandDFSsearchstrategiesfromnodeu(k=3).andedges.Atypicalsolutioninvolveshand-engineeringdomain-specificfeaturesbasedonexpertknowledge.Evenifonediscounts(a)Deepwalk(b)Node2vec(b)LineGeneralizedasMatrixFactorization121.JiezhongQiu,YuxiaoDong,HaoMa,JianLi,KuansanWang,JieTang:NetworkEmbeddingasMatrixFactorization:UnifyingDeepWalk,LINE,PTE,andnode2vec.WSDM2018:459-46712SummaryaboutNE➢Needtodefinesimilarityfunctionaccordingtodifferentapplications➢Shallowmodel➢FocusmoreonnetworktopologySometimeswithbettereffect➢Transductiveworkingmode➢…13➢DeeplearninghasachievedagreatsuccessSuchasCNN,LSTMetc➢DeepmodelforNetworkDataishardCNNsforfixed-sizeimages/grids….RNNsorword2vecfortext/sequences…GraphNeuralNetwork14shift-invariantGraphNeuralNetwork➢Gr...