Relational Inductive Biases, Deep Learning, and Graph Networks, Peter W. Battaglia, Jessica B. Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, Caglar Gulcehre, Francis Song, Andrew Ballard, Justin Gilmer, George Dahl, Ashish Vaswani, Kelsey Allen, Charles Nash, Victoria Langston, Chris Dyer, Nicolas Heess, Daan Wierstra, Pushmeet Kohli, Matt Botvinick, Oriol Vinyals, Yujia Li, Razvan Pascanu, Thomas P. Lillicrap, 2018arXiv preprint arXiv:1806.01261DOI: 10.48550/arXiv.1806.01261 - This paper defines the unified 'Graph Networks' framework, which encompasses message passing and formalizes its components.
How Powerful are Graph Neural Networks?, Keyulu Xu, Weihua Hu, Jure Leskovec, Yoshua Bengio, 2019International Conference on Learning Representations (ICLR)DOI: 10.48550/arXiv.1810.00826 - Introduces the Graph Isomorphism Network (GIN) and formally connects the expressive capacity of GNNs to the 1-Weisfeiler-Lehman test.
Weisfeiler and Lehman Go Neural: Higher-order Graph Neural Networks, Christopher Morris, Martin Ritzert, Matthias Fey, Stefan Faber, Hongzhi Wen, William Hamilton, Laurent Schonsberg, Michiharu Akimoto, Martin Gütlein, Daniel Kuehn, Karsten Borgwardt, 2019Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 33 (American Association for Artificial Intelligence (AAAI))DOI: 10.1609/aaai.v33i01.33013990 - Presents a theoretical framework for analyzing the expressive capacity of GNNs concerning the Weisfeiler-Lehman test, including higher-order variants.