L-SING

Learning local neighborhoods of non-Gaussian graphical models
Accepted at AAAI'25

Arxiv: https://arxiv.org/abs/2503.13899

AAAI: https://ojs.aaai.org/index.php/AAAI/article/view/34059/36214

Overview

L-SING (Localized Sparsity Identification for Non-Gaussian Distributions) solves a graph recovery problem: given $n$ i.i.d. from an (unspecified and possibly non-Gaussian) probability distribution, L-SING recovers the local neighborhood structure (local Markov properties) of each variable in the graph.

Authors

Sarah Liaw, Rebecca Morrison, Youssef Marzouk, Ricardo Baptista
Correspondence to: [email protected]

---

Experiments

This repo includes the codebase and experiment scripts used for our AAAI'25 submission. Note that unit tests are still being worked on right now.

We conducted three primary experiments to evaluate the performance of L-SING:

1. Butterfly Distribution (Non-Gaussian)

   • A synthetic experiment to show L-SING's ability to learn complex distributions with localized dependencies. We explain how the butterfly distribution is generated in the arxiv submission.

2. Gaussian Distribution

   • Demonstrates L-SING's compatibility with Gaussian distributions for benchmarking purposes.

3. Ovarian Cancer Dataset

   • From the curatedOvarianPackage in R.
   • The dataset was pre-processed following the methodology outlined by Shutta et al. (2022) for direct comparison between GLASSO and localized L-SING methods.

---

Usage

This repo includes examples of differently parameterized UMNNs to replicate results described in the AAAI'25 submission. For detailed experiment configurations (e.g., regularization values, specific UMNN architectures) in our paper, refer to the technical appendix in arxiv.

Running an Experiment

To execute an experiment, follow these steps:

1. Ensure Configuration Files Are Correct:

   • Modify the paths in the config.yaml file to point to the correct dataset files (training_file, validation_file, testing_file) and desired output directory (results_path).
   • Each experiment's script (e.g., run_butterfly.py) reads the paths directly from config.yaml.

2. Run the Experiment Script:

   • Navigate to the main project directory.
   • Run the experiment script using:

     python -m experiments.butterfly.run_butterfly

   • Replace butterfly with the appropriate experiment name for other experiments (e.g., gaussian, ovarian).

3. Results:

   • Results, including the precision matrix, model files, and plots, will be saved in a uniquely indexed folder within the results/ directory (e.g., results/BF0/ for the first experiment run).
   • The experiment configuration used for the run is also logged in log.txt within the results folder.

---

Changing Configurations

The config.yaml file allows you to customize parameters for each experiment:

• Dataset Paths:

  • Update training_file, validation_file, and testing_file to point to your generated .txt or .csv dataset files.

• UMNN Parameters:

  • Adjust hidden_layers, num_steps, and other model-related settings as wanted.

• Regularization and Training:

  • Change/Update regularizations, learning_rate, and max_epochs for alternative experimental setups.

---

Experiment Logs

After running an experiment, the corresponding log.txt file in the results directory will include:

• The configuration used for the run.
• Paths to the saved precision matrix, precision matrix, and models.

---

Citations

If you use this code, please cite the following references:

1. Wehenkel, A., & Louppe, G. (2019).
   Unconstrained Monotonic Neural Networks.
   In Wallach, H.; Larochelle, H.; Beygelzimer, A.; d'Alché-Buc, F.; Fox, E.; and Garnett, R., eds., Advances in Neural Information Processing Systems, Vol. 32. Curran Associates, Inc.

2. Shutta, K. H., Vito, R. D., Scholtens, D. M., & Balasubramanian, R. (2022).
   Gaussian Graphical Models with Applications to Omics Analyses.
   Statistics in Medicine, 41(25), 5150–5187.

---

Acknowledgments

The UMNN model implementation in this repository is adapted from the official source:
https://github.com/AWehenkel/UMNN.

For further inquiries or clarifications, please refer to our technical appendix or contact the authors.

Citations

If you using L-SING in an academic paper, please cite:

@misc{liaw2025learninglocalneighborhoodsnongaussian, title={Learning local neighborhoods of non-Gaussian graphical models: A measure transport approach}, author={Sarah Liaw and Rebecca Morrison and Youssef Marzouk and Ricardo Baptista}, year={2025}, eprint={2503.13899}, archivePrefix={arXiv}, primaryClass={cs.LG}, url={https://arxiv.org/abs/2503.13899}, }