Reproduce “Locality-Aware Graph-Rewiring in GNNs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Federico Barbero, Ameya Velingker, Amin Saberi, Michael Bronstein, Francesco Di Giovanni
Researchers from the University of Oxford, Google Research, and Stanford University developed Locality-Aware Sequential Rewiring (LASER) for Graph Neural Networks. This method effectively mitigates the over-squashing problem by adding sparse, locality-preserving connections through a sequence of rewired graphs. LASER consistently outperformed or matched state-of-the-art rewiring methods on long-range graph benchmarks and demonstrated superior scalability for large graphs.
Key insights
- The inherent trade-off between preserving locality and maintaining sparsity in graph rewiring can be overcome through a sequential, locality-aware approach.
- Gradually introducing connections based on original shortest-path distance, while selectively prioritizing pairs with low initial connectivity, effectively mitigates over-squashing without destroying local inductive biases.
- Treating different 'hop levels' of added edges as distinct relations within a multi-relational GNN framework allows the model to learn context-dependent message aggregations, enhancing expressivity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.