Research map / Graph learning / GNN architectures
On the Bottleneck of Graph Neural Networks and its Practical Implications
Uri Alon, Eran Yahav
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a fully-adjacent final layer | R-GIN · QM9 | average error: -39.54% | Improved |
| Add a fully-adjacent final layer | R-GAT · QM9 | average error: -44.58% | Improved |
| Add a fully-adjacent final layer | GGNN · QM9 | average error: -47.42% | Improved |
| Add a fully-adjacent final layer | GIN · NCI1 | error: 1.5% | Improved |
| Add a fully-adjacent final layer | GIN · ENZYMES | error: 8.1% over GIN-base; 2.5% over No Struct | Improved |
| Add a fully-adjacent final layer | Tuned GNNs · VarMisuse SeenProjectTest | accuracy: 88.4% | Improved |
| Add a fully-adjacent final layer | Tuned GNNs · VarMisuse UnseenProjectTest | accuracy: 83.8% | Improved |
| Double the hidden dimension instead of adding an FA layer (ablation) | GCN · QM9 | average error: -5.50% | Improved |
| Convert every layer to fully adjacent, removing reliance on original graph topology (ablation) | GCN · QM9 | average error: +1520% | Hurt |
| Add a second fully-adjacent layer (ablation) | GCN + FA · QM9 | average error: -43.30% versus base; +FA is -43.40% | No clear effect |
| Move the fully-adjacent layer to the penultimate position (ablation) | GCN · QM9 | average error: -45.2% | Improved |
| Use only 0.25 of the possible edges in the FA layer (ablation) | GCN · QM9 | average error: -8.4% | Improved |
Closest papers by meaning
- Understanding over-squashing and bottlenecks on graphs via curvature (ICLR 2022 · similarity 0.92)
- On Vanishing Gradients, Over-Smoothing, and Over-Squashing in GNNs: Bridging Recurrent and Graph Learning (NeurIPS 2025 · similarity 0.89)
- Get a GRIP, this will be a long TRIP: A Quantifiable Long-Range Framework for Verifying Over-squashing (NeurIPS 2026 · similarity 0.89)
- Effects of Dropout on Performance in Long-range Graph Learning Tasks (NeurIPS 2025 · similarity 0.89)
- Over-squashing in Spatiotemporal Graph Neural Networks (NeurIPS 2025 · similarity 0.88)
- Graph Neural Networks Do Not Always Oversmooth (NeurIPS 2024 · 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.
