Research map / Graph learning / GNN architectures
MuseGNN: Forming Scalable, Convergent GNN Layers that Minimize a Sampling-Based Energy
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MuseGNN, a scalable unfolded GNN whose layers minimize a sampling-based graph-regularized energy over offline-sampled subgraphs, with convergence guarantees and SOTA homogeneous results on OGB and IGB up to IGB-full (>1TB).
Problem
Unfolded GNNs (whose layers minimize a graph-regularized energy) are hard to scale to very large graphs, and existing scalable variants lack convergence guarantees or large-scale evaluation.
Approach
Build offline subgraph sampling directly into the architecture-inducing energy (per-subgraph embeddings tied via a shared online-averaged summary M with penalty gamma), derive proximal-gradient unfolded layers from it, and prove convergence (gamma=0 global bilevel; arbitrary gamma lower-level).
Key results
- MuseGNN reaches 61.74% accuracy on IGB-full, above FreshGNN (56.50%) and GAT (NS) (55.51%), the best reported homogeneous result on that leaderboard.
- MuseGNN achieves 66.82% on ogbn-papers100M and 81.23% on ogbn-products, outperforming GAS-based baselines and LazyGNN on the larger datasets.
- Ablation over gamma shows gamma=0 already gives good accuracy (66.29% on papers100M), with gamma>0 yielding modest gains; forward-pass energy converges within 8 unfolded iterations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase gamma from 0 to 3 (ablation) | MuseGNN · ogbn-papers100M | accuracy (%): 66.29 -> 66.82 | Helped |
| Increase gamma from 0 to 0.1 (ablation) | MuseGNN · ogbn-arxiv | accuracy (%): 72.51 -> 72.67 | Helped |
| Increase gamma from 0 to 0.5 (ablation) | MuseGNN · IGB-tiny | accuracy (%): 72.66 -> 73.42 | Helped |
| Increase gamma from 0 to 0.5 (ablation) | MuseGNN · IGB-medium | accuracy (%): 75.18 -> 75.83 | Helped |
| Increase gamma from 0 to 0.5 (ablation) | MuseGNN · ogbn-products | accuracy (%): 80.42 -> 81.23 | Helped |
| Increase gamma beyond 0.1 (ablation) | MuseGNN · ogbn-arxiv | accuracy (%): 72.67 -> 72.08 | Hurt |
| Use online neighbor sampling instead of offline ShadowKHop (ablation) | MuseGNN · ogbn-papers100M | accuracy (%): 66.82 -> 66.11 | Hurt |
| Use offline ShadowKHop instead of online neighbor sampling (ablation) | GCN · ogbn-papers100M | accuracy (%): 65.83 -> 64.89 | Hurt |
| Use offline ShadowKHop instead of online neighbor sampling (ablation) | GAT · ogbn-papers100M | accuracy (%): 66.28 -> 63.84 | Hurt |
| Use offline ShadowKHop instead of online neighbor sampling (ablation) | SAGE · ogbn-papers100M | accuracy (%): 66.2 -> 65.57 | Hurt |
| Use MuseGNN instead of the strongest listed baseline | Node classification models · ogbn-arxiv, IGB-tiny, ogbn-products, ogbn-papers100M, MAG240M, IGB-full | accuracy (%): 72.74 vs 72.50; 73.37 vs 73.42; 81.21 vs 81.23; 66.28 vs 66.82; 67.15 vs 67.67; 56.50 vs 61.74 | Mixed |
Surprising findings
- Baseline GNNs (GCN, SAGE, GAT) perform worse when trained on the same offline ShadowKHop samples than with online neighbor sampling, so the gain is not from the sampler.
Limitations
- Convergence guarantee for the full bilevel process holds only for gamma=0 with restrictive assumptions (linear f, zero zeta, identity g); arbitrary gamma only has lower-level convergence.
- Hyperparameters for the largest datasets were set by heuristics rather than grid search, and LazyGNN/USP results are partly taken from papers.
Experimental setup
Datasets: ogbn-arxiv IGB-tiny ogbn-products IGB-medium ogbn-papers100M MAG240M IGB-full
Benchmarks: ogbn-arxiv IGB-tiny ogbn-products ogbn-papers100M MAG240M IGB-full
Metrics: test accuracy epoch time
Baselines: GCN (NS) SAGE (NS) GAT (NS) SGFormer MariusGNN FreshGNN GCN (GAS) SAGE (GAS) GAT (GAS) UGNN (FG) LazyGNN USP
Builds on: unfolded GNN energy framework proximal gradient descent ShadowKHop GNNAutoScale USP
Closest papers by meaning
- How Graph Neural Networks Learn: Lessons from Training Dynamics (ICML 2024 · similarity 0.88)
- Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence (ICML 2025 · similarity 0.88)
- On Vanishing Gradients, Over-Smoothing, and Over-Squashing in GNNs: Bridging Recurrent and Graph Learning (NeurIPS 2025 · similarity 0.88)
- Classic GNNs are Strong Baselines: Reassessing GNNs for Node Classification (NeurIPS 2024 · similarity 0.88)
- Making Classic GNNs Strong Baselines Across Varying Homophily: A Smoothness–Generalization Perspective (NeurIPS 2025 · similarity 0.88)
- Generative Graph Pattern Machine (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.