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..
Unsupervised Learning for Optimal Transport plan prediction between unbalanced graphs
Authors: Sonia Mazelet, Rémi Flamary, Bertrand Thirion
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate ULOT on synthetic stochastic block model (SBM) graphs and on real cortical surface data obtained from f MRI. ULOT predicts transport plans with competitive loss up to two orders of magnitude faster than classical solvers. Furthermore, the predicted plan can be used as a warm start for classical solvers to accelerate their convergence. We show in our experiments, on simulated and real life data, that our method outperforms classical solvers in terms of speed by two orders of magnitude while providing OT plans with competitive loss. |
| Researcher Affiliation | Academia | Sonia Mazelet CMAP, Ecole Polytechnique Palaiseau, France EMAIL Rémi Flamary CMAP, Ecole Polytechnique Palaiseau, France EMAIL Bertrand Thirion Mind, Inria-Saclay Palaiseau, France EMAIL |
| Pseudocode | No | The paper describes the architecture of ULOT in detail, including its components like node embedding with cross attention and transport plan prediction, but it does not present this information in a structured pseudocode or algorithm block. Figure 1 provides a high-level diagram. |
| Open Source Code | Yes | Our code is available at https://github.com/smazelet/ULOT |
| Open Datasets | Yes | We now evaluate ULOT on the Individual Brain Charting (IBC) dataset [23] which is a dataset of functional MRI activations on brain surfaces. [...] The data is public. |
| Dataset Splits | Yes | The final dataset is constructed from all possible graph pairs, which we randomly split into 60% training, 20% validation, and 20% test sets. |
| Hardware Specification | Yes | We trained our network on an NVIDIA V100 GPU for 100 hours on the IBC dataset and a few hours on the smaller simulated dataset. |
| Software Dependencies | No | Hyperparameter optimization was performed on a subset of the training data using the Optuna library [1]. ... We will also share the pre-trained model weights for both datasets. ... We find in figure 7 (right) that even though ULOT makes errors, it is up to 100 times faster than classical solvers and 10 times faster than Sinkhorn for a smaller error. This computational gain on graphs of size 1000 is very important as the solvers have a cubic time complexity with respect to the number of nodes, while ULOT has a quadratic time complexity as shown in Figure 8 (left) where computation time is plotted against the number of nodes. |
| Experiment Setup | Yes | Table 1: ULOT hyperparameters Hyperparameter Simulated dataset IBC dataset Learning rate 0.001 0.0001 Batch size 256 64 Optimizer Adam Adam Number of node embedding layers N 5 3 Embedding dimension for α 10 10 Node embedding layer final out dimension 256 256 MLP hidden dimension 64 256 GCN hidden dimension 16 128 Temperature value a 3 5 |