1-3 图形神经网络中的超越同伦:当前限制、有效设计和对鲁棒性的影响.pdf

1、Beyond Homophily in GNNs:Current Limitations,Effective Designs,and Impacts on RobustnessJiong ZhuPh.D.StudentUniversity of MichiganJoint work with:Danai Koutra,Yujun Yan,Lingxiao Zhao,Mark Heimann,Leman Akoglu,Ryan Rossi,Junchen Jin,Donald Loveland,Michael Schaub,This Talk2 Generalizing Graph Neural

2、 Networks Beyond Homophily Beyond Homophily in Graph Neural Networks:Current Limitations and Effective Designs.NeurIPS 2020.Graph Neural Networks with Heterophily.AAAI 2021.Relationship between Heterophily and Robustness of Graph Neural Networks How does Heterophily Impact Robustness of Graph Neural

3、 Networks?Theoretical Connections and Practical Implications.KDD 2022.This Talk3 Generalizing Graph Neural Networks Beyond Homophily Beyond Homophily in Graph Neural Networks:Current Limitations and Effective Designs.NeurIPS 2020.Graph Neural Networks with Heterophily.AAAI 2021.Relationship between

4、Heterophily and Robustness of Graph Neural Networks How does Heterophily Impact Robustness of Graph Neural Networks?Theoretical Connections and Practical Implications.KDD 2022.Limitations&DesignsNode Classification4CyberSecurityBotUserBots?RecommendationSystemsGraph Neural Networks Many Graph Neural

5、 Network(GNN)models proposed recently5Cinput layerX1X2X3X4Foutput layerZ1Z2Z3Z4hiddenlayersY1Y41Our focus:Characterizing the representation power of GNNs beyond the homophily settings.Plot:Kipf+ICLR17However,most existing GNN models only look intographs with strong homophily(i.e.,where linked nodes

6、are similar)and ignore other possibilities.Graphs:Homophily and the Beyond6“Opposites Attract”Majority of linked nodes are differentNewman Networks18,Newman 04,Lee+arXiv18,Chau+ECML/PKDD06Friend network(e.g.,talkative/silent friends)Protein structures(wrt.amino acid types)E-commerce(wrt.fraudsters/a