The article ‘Cooperative Graph Neural Networks’ by authors from the University of Oxford proposes a new training method for graph neural networks. The authors view each node as a player that can choose to listen, broadcast, isolate or do both. This approach allows each node to determine its strategy based on its state, providing a more dynamic way of message-passing. The authors provide a theoretical and empirical analysis of this new message-passing scheme. This new approach aims to overcome limitations related to information flow in the traditional message-passing paradigm.
Publication date: 2 Oct 2023
Project Page: https://arxiv.org/abs/2310.01267v1
Paper: https://arxiv.org/pdf/2310.01267
