Research map / Graph learning / GNN architectures
Optimality of Message-Passing Architectures for Sparse Graphs
Aseem Baranwal, Kimon Fountoulakis, Aukosh Jagannath
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 |
|---|---|---|---|
| Use Architecture 1 (Bayes-optimal local classifier) instead of vanilla GCN or MLP | Architecture 1 · CSBM n=10000, d=4, binary | test accuracy | Improved |
| Set graph signal Γ=0 or Γ=1 (extreme graph signals) | Architecture 1 · CSBM n=10000, d=4 | test accuracy | No clear effect |
| Increase AMP-BP iterations from t=5 to t=20 (ablation) | AMP-BP · CSBM n=1000, γ=1, d=5 | accuracy: 0.753 -> 0.858 (Γ=0.3) | Improved |
| Increase AMP-BP iterations from t=5 to t=20 (ablation) | AMP-BP · CSBM n=1000, γ=1, d=500 | accuracy: 0.554 -> 0.848 (Γ=0.3) | Improved |
| Architecture 1 versus AMP-BP t=20 in low dimension with strong features | Architecture 1 · CSBM n=1000, γ=1, d=5 | accuracy: 0.870 vs 0.858 (Γ=0.3); 0.996 vs 0.952 (Γ=0.6) | Improved |
| Architecture 1 versus AMP-BP t=20 in high dimension with strong features | Architecture 1 · CSBM n=1000, γ=1, d=500 | accuracy: 0.995 vs 0.877 (Γ=0.4); 0.998 vs 0.940 (Γ=0.6) | Improved |
| Weaken feature signal from γ=1 to γ=0.2 in high dimension (d=500) | Architecture 1 vs AMP-BP t=20 · CSBM n=1000, d=500 | accuracy: 0.762 vs 0.804 (Γ=0.5); 0.788 vs 0.878 (Γ=0.6) | Hurt |
| Weaken feature signal to γ=0.2 in low dimension (d=5) | Architecture 1 vs AMP-BP t=20 · CSBM n=1000, γ=0.2, d=5 | accuracy: 0.554 vs 0.579 (Γ=0.3); 0.997 vs 0.880 (Γ=0.6) | Mixed |
| Random parameter initialization with gradient descent | Architecture 1 · CSBM binary, Γ=0.2 | parameter convergence to ansatz | Improved |
Closest papers by meaning
- Generalization and Representational Limits of Graph Neural Networks (ICML 2020 · similarity 0.88)
- NodeFormer: A Scalable Graph Structure Learning Transformer for Node Classification (NeurIPS 2022 · similarity 0.88)
- Graph Coarsening with Message-Passing Guarantees (NeurIPS 2024 · similarity 0.87)
- Almost Surely Asymptotically Constant Graph Neural Networks (NeurIPS 2024 · similarity 0.87)
- Fast Graph Sharpness-Aware Minimization for Enhancing and Accelerating Few-Shot Node Classification (NeurIPS 2024 · similarity 0.86)
- Learning Graph Representation via Graph Entropy Maximization (ICML 2024 · similarity 0.86)
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.
