基于图神经网络的知识图谱推理.pdf

1、基于知识图谱的图神经网络推理演讲人：张永祺第四范式算法科学家2023 背景介绍GNN for KGKG for GNN总结与展望背景介绍Knowledge graph(KG)Graph representation:=,.Entities:real world objects or abstract concepts.Relations:interactions between/among entities.Fact/triples:the basic unit in form of(head entity,relation,tail entity),.KG is a semantic gra

2、ph Semantic information Structural informationRepresentative applicationsKGQA:Personalized recommendation:Drug discovery:Event forecasting:Knowledge graph Knowledge Graph=Knowledge+GraphA directed multi-relational graph.A graph-structured representation.Whole graph/subgraph as input.Symmetric,invers

3、e,asymmetric,composition(A,spouse,B)(B,spouse,A)(A,older,B)(B,younger,A)(A,location,USA)(A,isBrotherOf,B)(B,isFatherOf,C)(A,isUncleOf,C)Triples/paths as input.Learning frameworkTargets:Preserve as much information on the original graph as possible.Generalize to the unseen triples/entities.Observed t

4、riple!:maximize scoreUnobserved triple:minimize score(Lebron,part_of,Lakers)(A.Davis,spouse_of,Britney)(Lakers,located_in,L.A.)(Savan Nah,lives_in,L.A.)(Lebron,part_of,Rockets)(A.Davis,spose_of,Lebron)(Wall Street,located_in,L.A.)(Savan Nah,son_of,L.A.)KGrepresentationstructurepredictorLossiterative

5、ly updateLearning problems Representation learning Knowledge graph reasoning,=?,#,#$#,$,performancecomplexityTransE2011TransH2014RotatE2019DistMult2015SimplE/CP2018QuatE2019AutoSF2020ConvE2018Interstellar 2020RESCAL2011RSN2019performancecomplexityDRUM 2019Neural LP 2017SimplE/CP 2018MINERVA 2018Deep

6、Path2017PathRanking2011interpretabilityComplEx2017CompGCN2019DPMPN 2020RED-GNN 2022GraIL2020Two classes of methodsreal world KGRepresentation learningKnowledge reasoningsparse&large&incompletecombinatorial optimizationDesign scoring function tomeasure triples.Learning on paths/subgraphs.GNN for KGKG