Survey on Graph Rewiring to Fix Over-Squashing and Over-Smoothing in GNNs
A new survey examines graph rewiring techniques in Graph Neural Networks (GNNs) to address over-squashing and over-smoothing. Over-squashing compresses information from distant nodes, while over-smoothing makes node representations indistinguishable after repeated propagation. Both issues degrade information flow and limit GNN performance. The survey reviews state-of-the-art rewiring methods, covering theoretical foundations, practical implementations, and performance trade-offs. It was published on arXiv on November 26, 2024, under the Computer Science > Machine Learning category.
Key facts
- Graph rewiring modifies graph topology to enhance information propagation in GNNs.
- Over-squashing compresses information from distant nodes.
- Over-smoothing makes node representations indistinguishable.
- Both phenomena stem from message passing and input topology interaction.
- Survey provides comprehensive review of rewiring approaches.
- Covers theoretical underpinnings, practical implementations, and performance trade-offs.
- Published on arXiv on November 26, 2024.
- Listed under Computer Science > Machine Learning.
Entities
Institutions
- arXiv