rflperry/sparse_shift

Overview

Conditional independence tools for causal learning under the sparse mechanism shift hypothesis.

Local installation

From a clean python environment (e.g. conda create -n test python=3.9),

git clone https://github.com/rflperry/sparse_shift.git
cd sparse_shift
pip install -e .

Running experiments and generating figures

First navigate and install necessary packages

cd experiments
pip install -r requirements.txt

Teaser figure

cd experiments

python teaser_sparse_oracle_pc.py

Runs the teaser experiment and generates the figure.

Simulations

cd experiments/main_simulations

Then run the following commands to generate results and the camera-ready figures.

Bivariate power:

python run_experiment.py --experiment bivariate_power --quick
python run_experiment.py --experiment bivariate_multiplic_power --quick

python plot_bivariate_identifiability.py

Oracle rates:

python run_experiment.py --experiment oracle_rates --quick
python run_experiment.py --experiment oracle_select_rates --quick

python plot_oracle_rates.py

Empirical comparison simulations:

python run_experiment.py --experiment pairwise_power --quick

python plot_empirical_power.py

Remove --quick and add --n_jobs -2 to run the full paper experiments. Note that this can take a long time.

Cytometry experiment

cd experiments/cytometry
python run_cytometry_experiment.py --quick

python analyze_pvalues.py

Remove --quick and add --n_jobs -2 to run the full paper experiments.