《2024-1-27真实复杂场景下的图神经网络-天津大学何东晓.pdf》由会员分享,可在线阅读,更多相关《2024-1-27真实复杂场景下的图神经网络-天津大学何东晓.pdf(55页珍藏版)》请在三个皮匠报告上搜索。
1、汇报人:何东晓汇报人:何东晓 天津大学天津大学 教授教授汇报时间:汇报时间:20232023年年1 1月月2727日日真实复杂场景下的图神经网络真实复杂场景下的图神经网络CatalogueAdversarial Representation Mechanism Learning for Network Embedding01Block Modeling-Guided Graph Convolutional Neural NetworksImproving Distinguishability of Class for Graph Neural NetworksContrastive Learn
2、ing Meets Homophily:Two Birds with One Stone020304Adversarial Representation Mechanism Learning for Network EmbeddingDongxiao He,et al.Adversarial Representation Mechanism Learning for Network Embedding,IEEE Transactions on Knowledge and Data Engineering(TKDE),2023,35(2):1200-1213.Introduction Graph
3、 representation learning:Graph representation learning aims to transform nodes on the graph into low-dimensionaldense vectors whilst still preserving the attribute features of nodes and structure features ofgraphs.表征学习图表征下游任务节点级别任务边级别任务图级别任务|d表征学习图表征下游任务节点级别任务边级别任务图级别任务34256798110=(,)|dIntroduction
4、Graph representation learning based on GCN:X=()1|()|Feature TransformationNeighborhood Aggregation313Neighborsof node 2(2)IntroductionGAN is inspired by the two-player game in game theory,which contains:A generator G(generating data that resemble real data).The generators goal is to foolt
5、he discriminator by generating data that are as similar to the real dataas possible.A discriminator D(distinguishing real data from generated data).The discriminators goalis to debunk the generator by discriminating between real data and generated data.GenerativeAdversarial Network(GAN):Introduction
6、ARGA:PreliminariesSymbolNotation=(,)an undirected,unweighted and attributed network =1,2,nodes=a set of edges ma set of node attribute=adjacency matrixThe objective of network embedding is to cast each of the nodes inthe network to a vector.Notations and the ProblemThe Approach-ArmGANAutoencoder wit
7、h mutual informationregularity(encoder,decoder and mutualinformation regularity)OverviewRepresentation mechanismdiscrimination(discriminator)Negative sample generator(negativeencoder,negative decoder and negativemutual information regularity)The Autoencoder with mutual information regularityTwo laye
8、r GCN encoderThe reconstructed loss(Structure only)Multual information regularity loss(Add the constraints of attribute information)Three-layer MLPWe randomly shuffle the rows of node attributes X to get thecorrupted,which is next encoded to get the corrupted embedding.The overall objective function
9、 of the autoencoder with mutual information regularityNegative sample generatorThe negative sample generator also uses the framework ofautoencoder with mutual information regularity but with adifferent optimization objective.In order to make the generated“fake”representationmechanism competitive,her
10、e we preserve the originaltopological structure but corrupt the node attributes,viarow-wise shuffling of real node attributes.NOT sharing parametersRepresentation Mechanisms Discrimination The core of our model is to adversarially learn therepresentation mechanism rather than the representationresul
11、t.The challenge is how to turn these two types ofrepresentation mechanism into recognizable inputs of therepresentation mechanism discriminator.Representation mechanism:Direct mapping representation mechanism(ArmGAN_d)Positive:Negative:Can be represented asMutual information representation mechanism
12、(ArmGAN_m)Positive:Negative:In essenceModel Training Discriminator TrainingGenerator TrainingArmGAN_d:ArmGAN_m:ORPositive generator training:Negative generator training:ArmGAN_d:ArmGAN_m:Generator training:ArmGAN_d:ArmGAN_m:OR The results of ArmGAN on 7 networks in classification,clustering and link
13、 prediction.ExperimentsNode classificationNode ClusteringLink Prediction(a):Representations are tightly interweaved.(b):Nodes are divided into different categories more clearly.ExperimentsVisualization We propose a new generative adversarial framework for networkembedding called ArmGAN,which uses ad
14、versarial learning strategy onthe representation mechanism rather than on embedding results so as tobetter.The new generative adversarial framework contains three players.Experimental results show that the new method significantly outperformsthe state-of-the-art methods including a typical GAN based
15、 method and amutual information based method.ConclusionDongxiao He,et al.Block Modeling-Guided Graph Convolutional Neural Networks,Proceedings of the ThirtySixth AAAI Conference on Artificial Intelligence(AAAI-22),2022,pp.4022-4029.Block Modeling-Guided Graph Convolutional Neural NetworksGraph Convo
16、lutional Neural NetworksEssence:propagate node attributes in neighbors guided by graph structure.Goal:encode nodes to embedding space by preserving network topology and attribute information.Is GCN Universal?Question:Can neighbor information represent a node?GCN workGCN NOT workHomophily Assumption:
17、GCN only works on homophily networks.most connections happen among nodes in the same or similar classesExisting Related Work Aggregating higher-order neighbors:H2GCN Zhu et al.,2020 MixHop Abu-El-Haija et al.,2019 Passing signed messages:GGCN Yan et al.,2021 GPR-GNN Chien et al.,2019 damage network
18、topologyfail to define optimal aggregating mechanismautomatically learn corresponding aggregation rulesfor neighbors of different classesA BETTER WAYHomophily Ratio&Block Matrix Block Matrix:measure the connected possibility of nodes in any two classes.Challenge 1:how to derive the block matrix in G
19、CN without all known labels.Challenge 2:how to design aggregation mechanism based on block matrix.block matrix depicts the heterophilic property of the network in heterophilic situations.Overview BM-GCNBlock Similarity Matrix1)Pre-train MLP for soft labels2)Calculate block matrix 3)Calculate block s
20、imilarity matrix obtain predicted soft labels via node attributes train MLP in a semi-supervised way make full use of existing known labels depicts the connecting pattern between classes depicts the similarity of connecting patternsbetween classes The more similar the two classes,the greater the val
21、ue of the corresponding element in Illustration on Why Block Modeling EffectiveHomophilicCora HeterophilicChameleon Block-Guided Graph Convolution ProcessSoft label vector:1234Soft label vector:1234The probability that belongs to 2-th classThe probability that belongs to 3-th class2,323The probabili
22、ty that(,)belongs to(2-3)class-combinationInformation propagating ratio between 2-th class and 3-th class 232,3Under the(2-3)class-combination,How many message should be propagated between and Consider all kinds of class-combinations,information propagating ratio between and Block-Guided Graph Convo
23、lution Process Information propagating ratio for two nodes Information propagating ratio for all node pairs(in form of matrix)A refined topology matrix based on New graph convolutional layer Simi-supervised Model Optimization(with fine-tuning MLP)ExperimentsHeterophilic NetworksHomophilic NetworksEx
24、perimentsConclusion Propose a new framework to make GCN applicable to both homophilic and heterophilic networks.Introduce block modeling technology to solve the problem of Homophily Assumption.Propose a novel design of block similarity matrix to enable block modeling technology to guide GCN to achie
25、ve classified aggregation.Dongxiao He,et al.Improving Distinguishability of Class for Graph Neural Networks.In Proceedings of the 38th AAAI Conference on Artificial Intelligence(AAAI-24).Improving Distinguishability of Class for Graph Neural NetworksDeep Graph Neural Networks(Deep GNNs)Goal:Learn co
26、mplex representations at different levels of abstraction from larger neighborhoods.Usual Framework:Stack multiple aggregations and transformations in a layer-wise fashion to construct a deep graph neural networks.Input Graph1-th Layer2-nd Layer3-th LayerFigure 1.The framework of deep graph neural ne
27、tworks.Does Deep GNNs Work?Phenomenon:How about the performance of deep GNNs?30%40%50%60%70%80%90%2 layers4 layers6 layers8 layers10 layersACC Essence:How does node representations change in deep GNNs?L layers?Figure 2.The performance of GCN on Cora datasets.Existing Related Work Over-smoothing prob
28、lemNoderepresentationsfromdifferentclasses are too similar or indistinguishable.JK-Net Xu et al.,2018GCNII Chen et al.,2020AERO-GNN Lee et al.,2023IndistinguishableihjhDistinguishable?Ignore whether different features within the same node representation can be distinguished from each other.Motivatio
29、n Feature-wise Metric:Local Distinguishability of Class(LDC)Global Distinguishability of Class(GDC)Node-wise Metric:()()()2,1|ijlllijv vVGSzzN=Global Smoothness(GS)(a)Cora(b)Citeseer(c)Texas(d)CornellFigure 3.As GCN layers stacking,The trends of ACC,GDC and GS.Motivation Howtoadaptivelylearnnoderepr
30、esentationswithhighdistinguishability of class for each node?(a)2-nd GCN layers(b)10-th GCN layers(c)12-th GCN layersFigure 4.The distribution of LDC on Cora dataset under the 2-nd,10-th and 12-th GCN layers,respectively.Overview Disc-GNNInter-layer FilteringLDC-based gating mechanismThe filtering d
31、iagonal matrixLDC-based inter-layer filtering where LDC(1)andLDC()correspond to the final representationgenerated in(l-1)-th layers and the representation generated by themessage passing mechanism in l-th layers,respectively.Initial Compensation&Global OptimizationGDC-based global optimizationThe co
32、mpensating diagonal matrixLDC-based initial compensationExperimentsHomophilic DatasetsHeterophilic DatasetsExperimentsExperimentsDepth analysisAblation Study(a)Cora(b)Texas(a)Cora(b)TexasConclusion Design two new metrics named Global Distinguishability of Class(GDC)and Local Distinguishability of Cl
33、ass(LDC)to evaluate the changes in node representations as GNN layers deepen.Propose Graph Neural Network guided by Distinguishability of class(Disc-GNN)can ensure that the learned node representations have a certain ability on distinguishing for different classes.Experimental results on both node c
34、lassification and node clustering tasks demonstrate the effectiveness of our proposed Disc-GNN.Dongxiao He,et al.Contrastive Learning Meets Homophily:Two Birds with One Stone,In International Conference on Machine Learning(ICML-23),2023,pp.12775-12789.Contrastive Learning Meets Homophily:Two Birds w
35、ith One StoneGraph Self-Supervised Contrastive Learning44It is known that Graph Neural Networks can perform well inlots of scenarios.Although most of them are semi-supervised,they need manually annotated and expensive data to train.Soit s necessary to explore self-supervised methods to train themode
36、l.Inheriting Computer Vision,existing methods almost augmentviews for positive or negative sampling.Then they maximizethemutualinformationbetweenthepositivepairsandminimize the mutual information between the negative pairs.Data Augmentation45The sampling method under a label consistent assumption,wh
37、ich means the augemented positive pairsshould have same semantic information.This is too strict for graph.Because most augmentation methods are based on random distribution.GRACE,BGRL:random edge dropping,random node masking.GCA:unimportant nodes and edges are randomly perturbed with higher ratios a
38、ccording to theircentrality rankingAD-GCL:Using deeplearning,the augmented views are generated based on the idea of max-minThese methods cant guarantee the abstracted semantics of nodes remain unchanged.So we are motivated to explore graph self-supervised methods without augmentation.label consisten
39、t assumptionlabel consistent assumptionA Crossroads of Challenges46Most universal main stream GNNs work under homophily assumption which means the linkednodes always have same semantics.Message passing on non-homophily edges leads to fusing ofinformation from different classes which degrades the dow
40、nstream tasks performance.For graph self-supervised contrastive learning without augmentation,how to select positivesamples that satisfy label consistent assumption is a challenge.In almost all current GSSL works,Graph Neural Networks(GNN)are used as encoders.How to integrate the sampling problem of
41、 graph contrastive learning and the homophily problem of graph neural networks,and how to solve them simultaneously?Motivation ExperimentObservation As the graphs rate t for node pairs increases,both GCN and GRACE significantly improve upon the traditional methods with the original topology.One-hop
42、neighbor set is a rich source ofpositive samplesOverview-NeCoHomophily Discrimination491.Parameterize the Homophily Discrimination2.Gumbel-Max TrickWe employ the Gumbel-Max trick to turn this continuous probability of node pairs into categorical samplings.represent the label of node and .Note that t
43、he learned stronger homophily structure is used in boththe message passing in GNNs and positive samplings in GCL.where andLoss Function50Positive SamplesNegative SamplesEdge Constraint TermThe edge constraint term is addedso that the model will not simplydelete all edges to reduce the loss.Experimen
44、ts51*The Supervised GCN was trained using 10%of the labelsConclusion We integrated the positive sampling strategy of GCL and the homophily discrimination of GNNs in the same framework and developed a new idea of solving them simultaneously We have proposed to extend the range of positive samplings t
45、o node neighbor sets.It allows us to develop a new paradigm of contrastive learning to avoid creating extra views in the traditional contrastive models and to eliminate data augmentation completely To address the inter-class node pairs in neighbor sets,we proposed a parametric homophily discriminati
46、on module.PublicationsDongxiao He,Shuwei Liu,Meng Ge,Zhizhi Yu*,Guangquan Xu,Zhiyong Feng.Improving Distinguishability of Class for Graph Neural Networks,In Proceedings of the 38th AAAI Conference on Artificial Intelligence(AAAI-24),2024.(CCF A 类会议长文)Dongxiao He,Jitao Zhao,Cuiying Huo,Yongqi Huang,Y
47、uxiao Huang&Zhiyong Feng,A New Mechanism for Eliminating Implicit Conflict in Graph Contrastive Learning,Proceedings of the Thirty-Eight AAAI Conference on Artificial Intelligence(AAAI-24),2024.(CCF A类会议长文)Yiqi Dong,Dongxiao He,Xiaobao Wang,Youzhu Jin,Meng Ge,Carl Yang&Di Jin.Unveiling Implicit Dece
48、ptive Patterns in Multi-modal Fake News via Neuro-Symbolic Reasoning,Proceedings of the 38th AAAI Conference on Artificial Intelligence(AAAI-24),2024.(CCF A类会议长文)Luzhi Wang,Di Jin,He Zhang,Yixin Liu,Dongxiao He,Wenjie Wang,Shirui Pan&Tat-Seng Chua,GOODAT:Towards Test-time Graph Out-of-Distribution D
49、etection,Proceedings of the 38th AAAI Conference on Artificial Intelligence(AAAI-24),2024.(CCF A类会议长文)Dongxiao He,Jitao Zhao,Rui Guo,Di Jin,Zhiyong Feng,Yuxiao Huang,Weixiong Zhang,Zhen Wang,Contrastive Learning Meets Homophily:Two Birds with One Stone,(ICML-23),2023:12775-12789(CCF-A类会议长文)Dongxiao
50、He,Tao Wang,Lu Zhai,Di Jin,Liang Yang,Yuxiao Huang,Zhiyong Feng&Philip S.Yu,Adversarial Representation Mechanism Learning for Network Embedding,IEEE Transactions on Knowledge and Data Engineering(TKDE),2023,35(2):1200-1213.(CCF-A类期刊)Di Jin,Zhizhi Yu,Dongxiao He*,Carl Yang,Philip S.Yu&Jiawei Han,GCN
51、for HIN via Implicit Utilization of Attention and Meta-paths,IEEE Transactions on Knowledge and Data Engineering(TKDE),2023,35(4):3925-3937.(CCF-A类期刊)Di Jin,Zhizhi Yu,Pengfei Jiao*,Shirui Pan,Dongxiao He*,Jia Wu,Philip S.Yu&Weixiong Zhang,A Survey of Community Detection Approaches:From Statistical M
52、odeling to Deep Learning,IEEE Transactions on Knowledge and Data Engineering(TKDE),2023,35(2):1149-1170.(CCF-A类期刊)Liang Yang,Qiuliang Zhang,Chuan Wang,Xiaochun Cao,Bingxin Niu,Runjie Shi,Wenmiao Zhou,Dongxiao He*,Yuanfang Guo*&Zhen Wang,Graph Neural Network without Propagation,The Web Conference(WWW
53、-23),2023:469-477.(CCF-A类会议长文)Cuiying Huo,Di Jin,Yawen Li,Dongxiao He*,Yu-Bin Yang&Lingfei Wu,T2-GNN:Graph Neural Networks for Graphs with Incomplete Features and Structure via Teacher-Student Distillation,Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence(AAAI-23),2023:433
54、9-4346.(CCF-A类会议长文)Zhizhi Yu,Di Jin,Ziyang Liu,Dongxiao He*,Xiao Wang,Hanghang Tong&Jiawei Han,Embedding Text-Rich Graph Neural Networks with Sequence and Topical Semantic Structures,Knowledge and Information Systems(KAIS),2023,65(2):613-640.(CCF-B类期刊)主要合作者54于智郅天津大学助理研究员王晓宝天津大学助理研究员何东晓天津大学教授金弟天津大学教授个人主页http:/