Understanding and Extending Subgraph GNNs by Rethinking Their Symmetries

Explore a comprehensive lecture on subgraph Graph Neural Networks (GNNs) and their symmetries. Delve into the design space of Subgraph GNN architectures, their theoretical properties, and expressive power. Learn about node-based subgraph selection policies, the link between Subgraph GNNs and Invariant Graph Networks (IGNs), and a novel family of message-passing layers. Discover the SUN architecture, which unifies previous approaches while offering improved empirical performance. Cover topics including message passing neural networks, expressive GNNs, subgraphs and components, node-based Subgraph GNNs, IGNs, expressiveness landscape, and experimental results in molecular modeling. Gain insights from Q&A sessions and explore potential applications in graph-based machine learning tasks.