Sample-Efficient L0-L2 Constrained Structure Learning of Sparse Ising Models

This repo contains code for reproducing the results in the paper Sample-Efficient L0-L2 Constrained Structure Learning of Sparse Ising Models published at the 35th AAAI Conference on Artificial Intelligence (AAAI 2021).

Getting started

Dependencies can be installed via

pip install -r requirements.txt
python setup.py develop

Ising model

This is the main folder. It contains the following files:

• data.py generates data for Ising models with the topologies described in the paper, without or with the use of a Gibbs-sampler
• test_data.py test the data generation code
• l1_logreg.py implements the L1 logistic regression estimator
• l1_constrained_logreg.py implements the L1-constrained logistic regression estimator
• l1_ise.py implements the L1 interaction screening estimator
• l0_l2constrained_logreg.py implements the proposed L0-L2 logistic regression estimator
• l0_l2constrained_ise.py implements the proposed L0-L2 interaction screening estimator

We refer to the main paper for a presentation of the different algorithms.

Scripts

This folder contains the bash scripts used to generate the results in the main paper for Examples 1 and 2.

Each bash script calls the compare_ising_models.py python script.

test_compare_ising_models_P16-ex1_small.sh tests the implementation for a smaller number of independent problems (4 instead of 30).

Notebooks

This folder contains a notebook which displays the simulation results as graphs which are similar to the ones in the main paper.

The default code assumes that the script compare_ising_models_P16-ex1.sh has been run.