Research map / Graph learning / GNN architectures
Graph Neural Networks with Local Graph Parameters
Pablo Barcelo, Floris Geerts, Juan L Reutter, Maksimilian Ryschkov
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 cycle homomorphism counts C3–C10 | GNN models · ZINC | MAE: GAT: 0.47±0.02 -> 0.22±0.01; GCN: 0.35±0.01 -> 0.20±0.01; GraphSage: 0.44±0.01 -> 0.24±0.01; MoNet: 0.25±0.01 -> 0.19±0. | Improved |
| Use cycle homomorphism counts instead of cycle subgraph-isomorphism counts (ablation) | GNN models · ZINC | MAE: GAT: 0.24±0.01 vs 0.22±0.01; GCN: 0.22±0.01 vs 0.20±0.01; GraphSage: 0.24±0.01 vs 0.24±0.01; MoNet: 0.16±0.01 vs 0.19±0. | Mixed |
| Add K3, K4, and K5 homomorphism counts | GAT, GCN, GraphSage, and MoNet · PATTERN | Accuracy: GAT: 78.83 ± 0.60 -> 85.50 ± 0.23; GCN: 71.42 ± 1,38 -> 82.49 ± 0.48; GraphSage: 70.78 ± 0,19 -> 85,85 ± 0.15; MoNet: 85 | Improved |
| Add no homomorphism-count patterns (ablation) | GatedGCN · PATTERN | Accuracy: 86.15 ± 0.08 -> 86.15 ± 0.08 | No clear effect |
| Add the selected clique homomorphism counts | GAT, GCN, GraphSage, MoNet, and GatedGCN · COLLAB | Hits@50: GAT: 50.32±0.55 -> 52.87±0.87; GCN: 51.35±1.30 -> 54.60±1.01; GraphSage: 50.33±0.68 -> 51.39±1.23; MoNet: 49.81±1.56 -> | Improved |
| Use only C3 homomorphism counts (ablation) | GAT · ZINC | MAE: 0.47±0.02 -> 0.45±0.01 | Improved |
| Use only C4 homomorphism counts (ablation) | GAT · ZINC | MAE: 0.47±0.02 -> 0.34±0.02 | Improved |
| Use only C6 homomorphism counts (ablation) | GAT · ZINC | MAE: 0.47±0.02 -> 0.31±0.01 | Improved |
| Use C5 and C6 homomorphism counts (ablation) | GAT · ZINC | MAE: 0.47±0.02 -> 0.28±0.01 | Improved |
| Use homomorphism counts of C3 through C6 (ablation) | GAT · ZINC | MAE: 0.47±0.02 -> 0.23±0.01 | Improved |
| Extend cycle features from C3–C6 to C3–C10 (ablation) | GAT · ZINC | MAE: 0.23±0.01 -> 0.22±0.01 | Improved |
| Use all cycle counts rather than individual or smaller cycle sets (ablation) | GAT · ZINC | MAE: None: 0.47±0.02; {C3}: 0.45±0.01; {C4}: 0.34±0.02; {C6}: 0.31±0.01; {C5,C6}: 0.28±0.01; {C3,...,C6}: 0.23±0.01; {C3,..., | Mixed |
Closest papers by meaning
- A Theoretical Comparison of Graph Neural Network Extensions (ICML 2022 · similarity 0.92)
- A New Perspective on "How Graph Neural Networks Go Beyond Weisfeiler-Lehman?" (ICLR 2022 · similarity 0.91)
- Provably Powerful Graph Networks (NeurIPS 2019 · similarity 0.91)
- Generalization and Representational Limits of Graph Neural Networks (ICML 2020 · similarity 0.91)
- Analyzing the Expressive Power of Graph Neural Networks in a Spectral Perspective (ICLR 2021 · similarity 0.91)
- Graph Neural Networks Are More Than Filters: Revisiting and Benchmarking from A Spectral Perspective (ICLR 2025 · similarity 0.89)
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.
