Reproduce “Mitigating Over-Squashing in Graph Neural Networks by Spectrum-Preserving Sparsification” 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
Langzhang Liang, Fanchen Bu, Zixing Song, Zenglin Xu, Shirui Pan, Kijung Shin
Researchers at KAIST introduce GOKU, a Densification-Sparsification Rewiring (DSR) framework to mitigate over-squashing in Graph Neural Networks. This method reconstructs a denser latent graph and then sparsifies it, effectively enhancing connectivity while explicitly preserving the graph's Laplacian spectrum and maintaining sparsity.
Key insights
- The input graph can be conceptualized as a sparsified representation of a more optimally connected latent graph, allowing for reconstruction to identify and reintroduce critical "missing edges."
- Explicitly preserving the Laplacian spectrum during graph rewiring is crucial for retaining the graph's inherent structural and topological characteristics.
- A sequential Densification-Sparsification Rewiring (DSR) paradigm effectively balances enhancing connectivity with maintaining graph sparsity and spectral fidelity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.