Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Enforcing convex constraints in Graph Neural Networks
Authors: Ahmed Rashwan, Keith Briggs, Chris Budd, Lisa Kreusser
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate Proj Net on four classes of constrained optimisation problems: linear programming, two classes of non-convex quadratic programs, and radio transmit power optimization, demonstrating its effectiveness across diverse problem settings. |
| Researcher Affiliation | Collaboration | Ahmed Rashwan University of Bath Keith Briggs BT Research Chris Budd University of Bath Lisa Kreusser University of Bath and Monumo |
| Pseudocode | Yes | Algorithm 1 The linear CAD algorithm |
| Open Source Code | No | Answer: [No] Justification: While we are currently not able to provide code, we are hoping to eventually release both code and data. |
| Open Datasets | No | The problem instances used in our evaluation were randomly generated using different graph distributions, as outlined in Appendix B. |
| Dataset Splits | No | The paper describes how problem instances and constraints were randomly generated but does not specify how these generated data were split into training, validation, or test sets for model evaluation. For example, it states "The problem instances used in our evaluation were randomly generated using different graph distributions, as outlined in Appendix B." and "Random constraints where sampled as described in B, but without the offset u." but does not detail the splits used for training the Proj Net models. |
| Hardware Specification | Yes | All experiments were conducted on a single machine with an NVIDIA Ge Force RTX 4090 GPU and an AMD Ryzen 9 7950X3D CPU. |
| Software Dependencies | No | The paper mentions several software tools and libraries such as Pytorch [50], Tensor Flow [45], Gurobi [31], gurobipy, OR-Tools [51], CVXPY [20], NVIDIA's cuOpt solver, PIQP solver [56], and scipy.minimize [60]. However, specific version numbers for these software dependencies are not provided. |
| Experiment Setup | Yes | To encourage the unconstrained vector w in Proj Net to be closer to the feasible set, we add an additional penalty term ch ||w PCpwq||2 to the loss function during training. The hyperparameter ch provides a trade-off between speed and performance: increasing ch speeds up the CAD algorithm by moving w closer to C, but may reduced task performance. For linear programming, we trained Proj Net models with 8 message passing layers for GNNv θ and no SVC layers. For all other application problems we used 8 message passing layers and 3 SVC layers. |