Optimizing over trained GNNs via symmetry breaking
Authors: Shiqiang Zhang, Juan Campos, Christian Feldmann, David Walz, Frederik Sandfort, Miriam Mathea, Calvin Tsay, Ruth Misener
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To test our symmetry-breaking strategies and optimization formulations, we consider an application in molecular design. [...] Numerical results show the outstanding improvement of symmetry-breaking. |
| Researcher Affiliation | Collaboration | Shiqiang Zhang Imperial College London London, UK Juan S. Campos Imperial College London London, UK Christian Feldmann BASF SE Ludwigshafen, Germany David Walz BASF SE Ludwigshafen, Germany Frederik Sandfort BASF SE Ludwigshafen, Germany Miriam Mathea BASF SE Ludwigshafen, Germany Calvin Tsay Imperial College London London, UK Ruth Misener Imperial College London London, UK |
| Pseudocode | Yes | Algorithm 1 Indexing algorithm |
| Open Source Code | Yes | The code is available at https://github.com/cog-imperial/GNN_MIP_CAMD. |
| Open Datasets | Yes | We choose two datasets QM7 [104, 105] and QM9 [106, 107] from CAMD literature to test the proposed methods. |
| Dataset Splits | Yes | For QM7 dataset, [...] use the first 5000 elements to train and the last 822 elements to test. [...] For QM9 dataset, [...] the first 80000 elements are used to train and the last 28723 elements are used to test. |
| Hardware Specification | Yes | We performed all experiments on a 3.2 GHz Intel Core i7-8700 CPU with 16 GB memory. |
| Software Dependencies | Yes | GNNs are implemented and trained in Py G [95]. MIP formulations for GNNs and CAMD are implemented based on OMLT [79], and are optimized by Gurobi 10.0.1 [87] with default relative MIP optimality gap (i.e., 10^-4). |
| Experiment Setup | Yes | For each run, we first randomly shuffle the dataset, then use the first 5000 elements to train and the last 822 elements to test. Each model is trained for 100 iterations with learning rate 0.01 and batch size 64. |