1、关于网络嵌入和图卷积神经网络的一些思考崔 鹏清华大学Perspectives and Outlook on Network Embedding and GCN网络/图数据图是对于数据的一/通用、全面、复杂的表示形式网络无处不在社交网络生物网络金融网络物联网信息网络物流网络为什么网络很重要？我们很少只关心数据本身，而不关心数据之间的关联Reflected by relational subjects Decided by relational subjects TargetTargetImage CharacterizationSocial Capital网络数据对机器学习模型不友好G=(V,E

2、)LinksTopologyInapplicability of ML methodsNetwork DataFeature ExtractionPattern DiscoveryPipeline for network analysisNetwork ApplicationsLearnabilityLearning from NetworksNetwork EmbeddingGNNG=(V,E)G=(V)Vector SpacegenerateembedEasy to parallelCan apply classical ML methods网络嵌入(Network Embedding)网

3、络嵌入的目标Goal Support network inference in vector spaceReflect network structureMaintain network propertiesBACTransitivitypBasic idea:recursive definition of statespA simple example:PageRank图神经网络GNNF.Scarselli,et al.The graph neural network model.IEEE TNN,2009.定义在图拓扑上的学习框架pMain idea:pass messages betwe

4、en pairs of nodes&agglomeratepStacking multiple layers like standard CNNs:pState-of-the-art results on node classification图卷积神经网络GCNT.N.Kipf and M.Welling.Semi-supervised classification with graph convolutional networks.ICLR,2017.图神经网络GNN简史网络嵌入与图神经网络GraphFeatureNetwork EmbeddingGCNInputTask resultsM

5、odelOutputEmbeddingTask resultsFeatureTopology to VectorFusion of Topology and FeaturesUnsupervised vs.(Semi-)Supervised图卷积网络 v.网络嵌入p In some sense,they are different.p Graphs exist in mathematics.(Data Structure)p Mathematical structures used to model pairwise relations between objectsp Networks ex

6、ist in the real world.(Data)p Social networks,logistic networks,biology networks,transactionnetworks,etc.p A network can be represented by a graph.p A dataset that is not a network can also be represented by a graph.图卷积网络应用于自然语言处理pMany papers on BERT+GNN.pBERT is for retrieval.pIt creates an initial

7、 graph of relevant entities and the initial evidence.pGNN is for reasoning.pIt collects evidence(i.e.,old messages on the entities)and arrive at new conclusions(i.e.,new messages on the entities),by passing the messages around and aggregating them.Cognitive Graph for Multi-Hop Reading Comprehension

8、at Scale.Ding et al.,ACL 2019.Dynamically Fused Graph Network for Multi-hop Reasoning.Xiao et al.,ACL 2019.图卷积网络应用于计算机视觉pA popular trend in CV is to construct a graph during the learning process.pTo process multiple objects or parts in a scene,and to infer their relationships.pExample:Scene graphs.S

9、cene Graph Generation by Iterative Message Passing.Xu et al.,CVPR 2017.Image Generation from Scene Graphs.Johnson et al.,CVPR 2018.图卷积网络应用于符号推理pWe can view the process of symbolic reasoning as a directed acyclic graph.pMany recent efforts use GNNs to perform symbolic reasoning.Learning by Abstractio

10、n:The Neural State Machine.Hudson&Manning,2019.Can Graph Neural Networks Help Logic Reasoning?Zhang et al.,2019.Symbolic Graph Reasoning Meets Convolutions.Liang et al.,NeurIPS 2018.pStructural equation modeling,a form of causal modeling,tries to describe the relationships between the variables as a

11、 directed acyclic graph(DAG).pGNN can be used to represent a nonlinear structural equation and help find the DAG,after treating the adjacency matrix as parameters.图卷积网络应用于结构方程建模DAG-GNN:DAG Structure Learning with Graph Neural Networks.Yu et al.,ICML 2019.（大多数）图卷积网络方法的PipelinepCo-occurrence(neighborh

12、ood)网络嵌:拓扑向量化pHigh-order proximities网络嵌:拓扑向量化pCommunities网络嵌:拓扑向量化pHeterogeneous networks网络嵌:拓扑向量化（大多数）网络嵌入方法的PipelineLearning for Networks v.s.Learning via GraphsLearning for networksLearning Via GraphsNetwork EmbeddingGCN网络嵌入方法解决的核心问题Reducing representation dimensionality while preserving necessar

13、y topological structures and properties.Nodes&LinksNode NeighborhoodCommunityPair-wise ProximityHyper EdgesGlobal StructureNon-transitivityAsymmetric TransitivityDynamicUncertaintyHeterogeneityInterpretabilityTopology-driven图卷积神经网络方法解决的核心问题Fusing topology and features in the way of smoothing feature

14、s with the assistance of topology.Feature-driven如果问题是拓扑驱动的？p Since GCN is filtering features,it is inevitably feature-drivenp Structure only provides auxiliary information(e.g.for filtering/smoothing)p When feature plays the key role,GNN performs good p How about the contrary?p Synthesis data:stocha

15、stic block model+random featuresMethodResultsRandom10.0GCN18.31.1DeepWalk99.00.1网络嵌入 v.图神经网络There is no better one,but there is more proper one.反思：图神经网络是否真的是深度学习方法？p Recall GNN formulation:!#$=&!(,=*+,$/./0*+,$/.p How about removing the non-linear component:!#$=!(p Stacking multiple layers and add s

16、oftmax classification:12=3456789!:=3456789!($(:,$=3456789:!(Wu,Felix,et al.Simplifying graph convolutional networks.ICML,2019.High-order proximity30p This simplified GNN(SGC)shows remarkable results:Node classification Text Classification反思：图神经网络是否真的是深度学习方法？Wu,Felix,et al.Simplifying graph convolutional networks.ICML,2019.总结p Unsupervised vs.(Semi-)Supervisedp Learning for Networks vs.Learning via Graphsp Topology-driven vs.Feature-drivenp Both GCN and NE need to treat the counterpart as the baselinesTHANKS!THANKS!THANKS!