DRew: Dynamically Rewired Message Passing with Delay

Explore a conference talk on DRew (Dynamically Rewired Message Passing with Delay), a novel framework for improving Message Passing Neural Networks (MPNNs) in long-range interaction tasks. Delve into the challenges faced by MPNNs, including over-squashing, and discover how the proposed framework addresses these issues through layer-dependent rewiring and a delay mechanism. Learn about DRew instantiations of common MPNNs, understand why vDRew helps with over-smoothing, and examine experimental results on real-world datasets. Gain insights into future research directions and participate in a Q&A session to deepen your understanding of this innovative approach to graph neural networks.

- Intro
- Challenges with Message-Passing Neural Networks MPNNs
- Framework Proposal
- Dynamic Rewiring
- Introducing Delay
- DRew Instantiations of Common MPNNs
- Why Does vDRew Help with Over-Smoothing?
- Experiments
- Performance on Real-World Datasets
- Future Work
- Q+A