Research map / Graph learning / GNN architectures
Half-Hop: A graph upsampling approach for slowing down message passing
Mehdi Azabou, Venkataramana Ganesh, Shantanu Thakoor, Chi-Heng Lin, Lakshmi Sathidevi, Ran Liu, Michal Valko, Petar Veličković, Eva L Dyer
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 Half-Hop to GCN | GCN · Texas, Wisconsin, Actor, Squirrel, Chameleon, Cornell | accuracy: +16.75, +19.04, +7.80, +15.67, +21.80, +2.70 | Improved |
| Add Half-Hop to GAT | GAT · Texas, Wisconsin, Actor, Squirrel, Chameleon, Cornell | accuracy: +28.38, +34.12, +9.26, +9.58, +12.75, +10.81 | Improved |
| Add Half-Hop to GraphSAGE | GraphSAGE · Texas, Wisconsin, Actor, Squirrel, Chameleon, Cornell | accuracy: +3.51, +4.70, +2.59, +3.64, +4.25, -1.35 | Mixed |
| Add Half-Hop to GCN | GCN · Amazon Computers, Amazon Photos, Coauthor CS, WikiCS | accuracy: +0.70, +0.93, +0.65, +0.64 | Improved |
| Add Half-Hop to GraphSAGE | GraphSAGE · Amazon Computers, Amazon Photos, Coauthor CS, WikiCS | accuracy: +1.81, -0.48, +0.02, -0.86 | Mixed |
| Use Half-Hop as an additional augmentation alongside feature and edge dropout | BGRL · Amazon Computers, Amazon Photos, Coauthor CS, Coauthor Physics, WikiCS | accuracy: 91.02 vs 90.34; 93.88 vs 93.17; 93.61 vs 93.31; 95.75 vs 95.73; 80.76 vs 79.98 | Improved |
| Use Half-Hop as an additional augmentation alongside feature and edge dropout | GRACE · Amazon Computers, Amazon Photos, Coauthor CS, WikiCS | accuracy: 91.11 vs 89.53; 94.21 vs 92.78; 93.59 vs 91.12; 80.77 vs 80.14 | Improved |
| Train with Half-Hop and evaluate on the Half-Hopped graph | HH-GRACE · Amazon Photos, Coauthor CS | accuracy: 94.89 vs 92.78; 94.76 vs 91.12 | Improved |
| Remove the backward edge from target to slow node (ablation) | GCN + Half-Hop · Texas, Actor, Cornell | accuracy: 71.71 -> 68.8; 33.35 -> 32.17; 63.42 -> 57.66 | Hurt |
| Add a slow-node-to-source edge, allowing reverse path connectivity (ablation) | GCN + Half-Hop · Texas, Actor, Cornell | accuracy: 71.71 -> 58.47; 33.35 -> 31.93; 63.42 -> 42.16 | Hurt |
| Replace interpolated slow-node features with zeros (ablation) | GCN + Half-Hop · Texas, Actor, Cornell | accuracy: 72.88 -> 61.80; 33.39 -> 28.93; 63.33 -> 49.55 | Hurt |
| Replace interpolated slow-node features with uniform random features (ablation) | GCN + Half-Hop · Texas, Actor, Cornell | accuracy: 72.88 -> 53.33; 33.39 -> 24.70; 63.33 -> 37.48 | Hurt |
Closest papers by meaning
- Towards Dynamic Message Passing on Graphs (NeurIPS 2024 · similarity 0.88)
- Sign is Not a Remedy: Multiset-to-Multiset Message Passing for Learning on Heterophilic Graphs (ICML 2024 · similarity 0.88)
- Accurate and Scalable Graph Neural Networks via Message Invariance (ICLR 2025 · similarity 0.88)
- Probabilistically Rewired Message-Passing Neural Networks (ICLR 2024 · similarity 0.88)
- Multi-Track Message Passing: Tackling Oversmoothing and Oversquashing in Graph Learning via Preventing Heterophily Mixing (ICML 2024 · similarity 0.88)
- Probabilistic Graph Rewiring via Virtual Nodes (NeurIPS 2024 · similarity 0.87)
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.
