Research map / Graph learning / GNN architectures
Graph Positional Encoding via Random Feature Propagation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace l2 normalization with QR orthonormalization (ablation) | RFP with GINE on ZINC-12k · ZINC-12k | MAE: 0.1368 -> 0.1152 | Helped |
| Replace l2 normalization with QR orthonormalization (ablation) | RFP on OGBG-MOLHIV · OGBG-MOLHIV | ROC-AUC(%): 77.91 -> 79.83 | Helped |
| Replace l2 normalization with QR orthonormalization (ablation) | RFP on node classification (Squirrel, Texas, Cora, etc.) · Squirrel | accuracy (%): 48.89 -> 54.28 | Helped |
| Add DSS architecture processing multiple trajectories (ablation) | RFP - QR on ZINC-12k · ZINC-12k | MAE: 0.1152 -> 0.1117 | Helped |
| Add DSS architecture processing multiple trajectories (ablation) | RFP - QR on OGBG-MOLHIV · OGBG-MOLHIV | ROC-AUC(%): 79.83 -> 80.53 | Helped |
| Add learned propagation operator S_learn (ablation) | RFP - QR with DSS on ZINC-12k and OGBG-MOLHIV · ZINC-12k | MAE: 0.1117 -> 0.1106 (MAE); 80.53 -> 80.58 (ROC-AUC) | Mixed |
| Increase number of eigenvectors k from 4 to 16 (P=16) (ablation) | RFP - QR on ZINC-12k and OGBG-MOLHIV · ZINC-12k | MAE: 0.1305 (k=4) -> 0.1152 (k=16) | Helped |
| Increase number of propagation steps P from 1 to 16 (k=16) (ablation) | RFP - QR - A on node classification · Cora | accuracy (%): 86.93 (P=16,k=4) region; Table 9 shows accuracy rising with P | Helped |
| Add RFP trajectory (QR, adjacency) as PE on top of GCN | GCN on ogbn-arxiv · ogbn-arxiv | accuracy (%): 71.08 -> 71.97 | Helped |
| Add plain random node features as PE to GCN | GCN on ogbn-arxiv · ogbn-arxiv | accuracy (%): 71.08 -> 70.91 | Hurt |
| Use RFP (random init + propagation) instead of propagating raw input features (NFP) | QR propagation PE on node classification · Squirrel | accuracy (%): 46.85 (NFP) -> 54.28 (RFP) | Helped |
| Use RFP QR-based PE vs. Laplacian eigenvector PE on node classification | GCN/MLP backbones · Cora | accuracy (%): 86.31 (L EIGVECS 64) -> 87.99 (RFP-QR-A) | Helped |
Closest papers by meaning
- What functions can Graph Neural Networks compute on random graphs? The role of Positional Encoding (NeurIPS 2023 · similarity 0.91)
- What Improves the Generalization of Graph Transformers? A Theoretical Dive into the Self-attention and Positional Encoding (ICML 2024 · similarity 0.90)
- Learning Efficient Positional Encodings with Graph Neural Networks (ICLR 2025 · similarity 0.90)
- Sketch-Augmented Features Improve Learning Long-Range Dependencies in Graph Neural Networks (NeurIPS 2025 · similarity 0.88)
- Node Similarities under Random Projections: Limits and Pathological Cases (ICLR 2025 · similarity 0.88)
- A Manifold Perspective on the Statistical Generalization of Graph Neural Networks (ICML 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.