Research map / Graph learning / GNN architectures
Relational Pooling for Graph Representations
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 |
|---|---|---|---|
| Add relational pooling with permutation-sensitive node IDs to GIN | GIN · CSL graphs (10-class) | accuracy (%): 10.0 -> 37.6 | Helped |
| Add relational pooling to the Duvenaud model | Duvenaud et al. · HIV | ROC-AUC: 0.812 (0.014) -> 0.832 (0.013) | Helped |
| Add relational pooling to the Duvenaud model | Duvenaud et al. · MUV | ROC-AUC: 0.798 (0.025) -> 0.794 (0.025) | No clear effect |
| Add relational pooling to the Duvenaud model | Duvenaud et al. · Tox21 | ROC-AUC: 0.794 (0.010) -> 0.799 (0.006) | No clear effect |
| Restrict RP to induced subgraphs of size k=10 (ablation) | Duvenaud et al. · Tox21 | ROC-AUC: 0.794 (0.010) -> 0.687 (0.005) | Hurt |
| Restrict RP to induced subgraphs of size k=20 (ablation) | Duvenaud et al. · Tox21 | ROC-AUC: 0.794 (0.010) -> 0.755 (0.003) | Hurt |
| Vary the induced-subgraph size k (ablation) | Duvenaud et al. · Tox21 | ROC-AUC: 0.687 (0.005); 0.755 (0.003); 0.775 (0.011); 0.783 (0.007); 0.778 (0.007) | Mixed |
| Vary the induced-subgraph size k (ablation) | Duvenaud et al. · HIV | ROC-AUC: 0.812 (0.035); 0.813 (0.017); 0.829 (0.024); 0.807 (0.025); 0.818 (0.022) | No clear effect |
| Vary the induced-subgraph size k (ablation) | Duvenaud et al. · MUV | ROC-AUC: 0.773 (0.045); 0.777 (0.041); 0.776 (0.030); 0.776 (0.032); 0.768 (0.014) | No clear effect |
| Use a CNN with DFS orientation for the RP function | RP · HIV | ROC-AUC: 0.542 (0.004) | Hurt |
| Use a CNN with DFS orientation for the RP function | RP · MUV | ROC-AUC: 0.601 (0.042) | Hurt |
| Use a CNN with DFS orientation for the RP function | RP · Tox21 | ROC-AUC: 0.597 (0.006) | Hurt |
Closest papers by meaning
- From Relational Pooling to Subgraph GNNs: A Universal Framework for More Expressive Graph Neural Networks (ICML 2023 · similarity 0.93)
- Rethinking pooling in graph neural networks (NeurIPS 2020 · similarity 0.91)
- Hierarchical Graph Representation Learning with Pooling-Induced Substructures (NeurIPS 2026 · similarity 0.91)
- Geometry-Aware Edge Pooling for Graph Neural Networks (NeurIPS 2025 · similarity 0.90)
- Exploring Consistency in Graph Representations: from Graph Kernels to Graph Neural Networks (NeurIPS 2024 · similarity 0.88)
- Boosting Graph Pooling with Persistent Homology (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.