Research map / Graph learning / GNN architectures
Can Computational Reducibility Lead to Transferable Models for Graph Combinatorial Optimization?
Semih Cantürk, Thomas Sabourin, Frederik Wenkel, Michael Perlmutter, Guy Wolf
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 |
|---|---|---|---|
| replace local message passing (GCN, GIN, GATv2) with the GCON wavelet filter bank (ablation) | energy-based unsupervised solver · RB-small MVC | MVC size: 221.56 -> 211.69 | Improved |
| use Hamiltonian loss rather than quadratic first-term loss for MIS (ablation) | GCON · MIS on RB-small | independent set size: 15.5 -> 18.04 | Improved |
| fine-tune whole GNN with inverted output head for MIS to MVC transfer (ablation) | GCON · RB-small MIS->MVC | MVC size: 211.69 -> 211.56 | Improved |
| freeze GNN backbone and only retrain head for MIS<->MVC (ablation) | GCON · RB-small MIS and MVC | MIS size / MVC size: MIS 18.12 -> 17.68; MVC 211.69 -> 212.46 | Hurt |
| invert output head instead of resetting when backbone frozen (ablation) | GCON · RB-small MIS<->MVC | MIS size: 17.68 -> 17.69 | No clear effect |
| fine-tune full model (instead of frozen) when transferring MIS to MaxClique (ablation) | GCON · RB-small MaxClique, G features | clique size: 16.12 -> 16.55 | Improved |
| add 3-layer multi-head attention on frozen MIS backbone for MaxClique (ablation) | GCON · RB-small MaxClique | clique size: 16.12 -> 16.13 | No clear effect |
| fine-tune on complement graph with MIS-pretrained backbone (true reduction) (ablation) | GCON · RB-small MaxClique | clique size: 15.52 (frozen) -> 16.82 (fine-tuned) | Improved |
| pretrain on the other five tasks, then fine-tune for 20 epochs on K-coloring | GCON · BA-small leave-one-out | coloring violations: 61.92 -> 24.19 | Improved |
| pretrain on other tasks then fine-tune MDS | GCON · BA-small leave-one-out | MDS size: 35.57 -> 36.15 | No clear effect |
| pretrain on MDS, MIS, Coloring then fine-tune MaxClique, MaxCut, MVC for 20 epochs | GCON · BA-small | MVC size: 141.43 -> 139.70 | Improved |
| pretrain on MDS, MIS, Coloring then fine-tune MaxClique | GCON · BA-small | clique size: 4.31 -> 4.32 (full: 4.43) | No clear effect |
Closest papers by meaning
- Towards Pre-trained Graph Condensation via Optimal Transport (NeurIPS 2025 · similarity 0.88)
- GOAL: A Generalist Combinatorial Optimization Agent Learner (ICLR 2025 · similarity 0.88)
- Towards General Algorithm Discovery for Combinatorial Optimization: Learning Symbolic Branching Policy from Bipartite Graph (ICML 2024 · similarity 0.87)
- Covered Forest: Fine-grained generalization analysis of graph neural networks (ICML 2025 · similarity 0.87)
- Are Graph Neural Networks Optimal Approximation Algorithms? (NeurIPS 2024 · similarity 0.87)
- Generative Graph Pattern Machine (NeurIPS 2025 · 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.
