Catch my poster tomorrow at the NeurIPS MLSB Workshop! We present a simple (yet effective 😁) multimodal Transformer for molecules, supporting multiple 3D conformations & showing promise for transfer learning.

Interested in molecular representation learning? Let’s chat 👋!

5/6 How it works: Probabilistic sampling connects original nodes to virtual ones, enhancing connectivity without explicit pairwise computations. The result is a framework that achieves both higher WL expressiveness and efficiency in graph-based learning.

4/6 We demonstrate SOTA results on various benchmarks, effectively addressing over-squashing and under-reaching. IPR-MPNNs also surpass standard MPNNs in expressiveness, distinguishing complex graph structures—all while being faster and more memory-efficient than GTs. 🚀

3/6 Enter IPR-MPNNs: Our approach learns to rewire graphs probabilistically by adding virtual nodes. This eliminates the need for heuristics, making the method more flexible and task-adaptive, while maintaining computational efficiency. 🎯

1/6 We're excited to share our #NeurIPS2024 paper: Probabilistic Graph Rewiring via Virtual Nodes! It addresses key challenges in GNNs, such as over-squashing and under-reaching, while reducing reliance on heuristic rewiring. w/ Chendi Qian, @christophermorris.bsky.social @mniepert.bsky.social 🧵