图形神经网络:结合深度学习和符号增加（35页）——基于图神经网络认知的智能计算专场.pdf

1、Graph Neural Networks:Combing Deep Learning&Symbolic ReasoningLe SongPrincipal Engineer,Ant FinancialAssociate Professor,Georgia TechFundamental problem and challengeRepresent?Entire graphEach nodeApplyVectorRepresentationGraph TopologyNode attributeEdge attribute2!#!$!%!&!(#($(%(&GCN/GNN/MPN/Struct

2、ure2Vec)$(+)#(+)(+)%(+)(+)&(+)Obtain embedding viaiterative update algorithm:1.Initialize)-(+)=/0+!-,32.Iterate 4 times)-(5)/0#)-(57#)+0$9:!3($-.)Residual:!($)!($-.)+:;.!($-.)+Attention:!($):+;73 3!3($-.),3 3=1Gating:!($)1 E !($-.)+E :;.!($-.)+Many papers!#!$!%!&!(#($(%(&Different message passing sc

3、heme)#(+)1.Initialize)-.,1,22.Iterate 3 times3.Aggregate)-=5%9-)-(:),1)-.(;)=5#)-.(;#)+5$9-B)-(;#),(1,2)#(+)#$(+)Parameterized as neural networkDai,et al.ICML 16Obtain embedding viaiterative update algorithm:6Lecture-Style TutorialBenefit of GNN for Benefit of GNN for Graph Feature Extraction Algori

4、thmGraph Feature Extraction AlgorithmFraudulent account detectionAlipay:new accounts in a month:millions of nodes and edges.Fake account can increase system level risk?BadGooddeviceLeverage account activity+connectivity?8Liu,et al.CCS 17,CIKM17Fraudulent account patternFake accountNormal accountDevi

5、ce ConnectivityAccountActivity9Results215 years,trillion of eventsTime-varying dependency structureTemporal knowledge graph:What will happen next?Trivedi et al.ICML 201717Enemys friend is an enemyCAIROCAIROCROATIACROATIAMANCHESTERMANCHESTERPROTESTORPROTESTORNEW ZEALANDNEW ZEALANDSOMALIASOMALIATEHRAN

6、TEHRANSINGAPORESINGAPOREASSAULTASSAULT302015062015ASSAULT:062015CONSULTCONSULTCONSULTCONSULTCOOPERATECOOPERATE062015292015(predicted event)(predicted event)THREATENTHREATENPROVIDE AIDPROVIDE AID072015202015282015FIGHTFIGHTFIGHTFIGHT05201518COLOMBIAOTTAWAOTTAWANEW DELHINEW DELHIBELGIUMBELGIUMLIBYALIB