Research map / Graph learning / GNN architectures
Hybrid Models for Learning to Branch
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 |
|---|---|---|---|
| Use the hybrid FiLM model instead of the strongest CPU baseline | Branching strategy · Capacitated Facility Location small and medium; Set Covering small and medium; Maximum Independent Set small, medium, and big; Combinatorial Auctions medium and | solver running time: 8.73 vs 9.83; 63.75 vs 89.24; 2.13 vs 2.73; 15.71 vs 22.26; 24.67 vs 30.37; 136.42 vs 172.51; 531.70 vs 633.42; 131.45 v | Helped |
| Use the hybrid FiLM model instead of the strongest CPU baseline | Branching strategy · Set Covering big and Combinatorial Auctions small | solver running time: 217.02 vs 210.95; 52.96 vs 44.07 | Hurt |
| Train end-to-end rather than using a pretrained GNN (ablation) | FiLM · FiLM test sets | Top-1 accuracy: 44.12 ± 0.09 -> 44.31 ± 0.08; 65.78 ± 0.06 -> 66.33 ± 0.33; 53.16 ± 0.51 -> 53.23 ± 0.58; 50.00 ± 0.09 -> 50.16 ± 0.05 | Helped |
| Add knowledge distillation to end-to-end training (ablation) | FiLM · FiLM test sets | Top-1 accuracy: 44.31 ± 0.08 -> 44.10 ± 0.09; 66.33 ± 0.33 -> 66.60 ± 0.21; 53.23 ± 0.58 -> 53.08 ± 0.3; 50.16 ± 0.05 -> 50.31 ± 0.19 | Mixed |
| Add auxiliary representation-diversity training to e2e & KD (ablation) | FiLM · FiLM test sets | Top-1 accuracy: 44.10 ± 0.09 -> 44.56 ± 0.13; 66.60 ± 0.21 -> 66.85 ± 0.28; 53.08 ± 0.3 -> 53.68 ± 0.23; 50.31 ± 0.19 -> 50.37 ± 0.03 | Helped |
| Use exponential depth-dependent weighting rather than constant weighting (ablation) | MLP branching model · Combinatorial Auctions big | B&B nodes: 9678 -> 9793 | Hurt |
| Use linear depth-dependent weighting rather than constant weighting (ablation) | MLP branching model · Combinatorial Auctions big | B&B nodes: 9678 -> 9789 | Hurt |
| Use quadratic depth-dependent weighting rather than constant weighting (ablation) | MLP branching model · Combinatorial Auctions big | B&B nodes: 9678 -> 9561 | Helped |
| Use sigmoidal depth-dependent weighting rather than constant weighting (ablation) | MLP branching model · Combinatorial Auctions big | B&B nodes: 9678 -> 9534 | Helped |
| Use FiLM or CONCAT hybrids instead of a plain MLP (ablation) | Branching architecture · All four problem sets | Top-1 accuracy | Helped |
| Use HyperSVM or HyperSVM-FiLM instead of FiLM or CONCAT (ablation) | Branching architecture · All four problem sets | Top-1 accuracy | Hurt |
| Run the GNN on CPU instead of using the hybrid model | Branching strategy · CPU-only MILP solving | solver running time | Hurt |
Closest papers by meaning
- Rethinking the Capacity of Graph Neural Networks for Branching Strategy (NeurIPS 2024 · similarity 0.92)
- Expressive Power of Graph Neural Networks for (Mixed-Integer) Quadratic Programs (ICML 2025 · similarity 0.88)
- On the Power of Small-size Graph Neural Networks for Linear Programming (NeurIPS 2024 · similarity 0.87)
- DOGE-Train: Discrete Optimization on GPU with End-to-End Training (AAAI 2024 · similarity 0.87)
- BIPNN: Learning to Solve Binary Integer Programming via Hypergraph Neural Networks (NeurIPS 2025 · similarity 0.87)
- TINED: GNNs-to-MLPs by Teacher Injection and Dirichlet Energy Distillation (ICML 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.