Learning DAGs from Data with Few Root Causes

Description

This is the implementation of our method and experiments of our paper accepted in NeurIPS 2023.

If you find this repository useful please cite:

@inproceedings{misiakos2023learning,
      title={Learning DAGs from Data with Few Root Causes}, 
      author={Panagiotis Misiakos and Chris Wendler and Markus Püschel},
      booktitle = {Advances in Neural Information Processing Systems},
      year={2023},
}

SparseRC

Our method is called SparseRC (from sparser root causes). To use it you can simply install Python with the required packages in sparserc_requirements.txt

A demonstration of our DAG learning method is shown in SparseRC_demo.ipynb

Method description

We consider the formulation of linear SEMs as a linear transformation of the input (defined as root causes) that gives the data as output.

Motivated by this new formulation we propose the novel DAG learning method that assumes sparsity in the input side - namely few root causes.

Full experiment reproducibility

Current version of the code was run successfully in Windows 11 with python 3.9.7.

Environment and packages

Setup the python environment:

# create a new conda environment
conda create -n sparserc python=3.9.7
conda activate sparserc

# install all requirements with pip
pip install -r requirements.txt

Special care must be taken for installing the pydot package. We suggest installation using conda.

# install pydot with conda
conda install pydot

Other repos

Then manually import the following repositories.

# install py-causal repo
pip install git+https://github.com/bd2kccd/py-causal

# clone the following repositories
cd neurips_experiments/methods/
git clone https://github.com/xunzheng/notears
git clone https://github.com/cdt15/lingam
git clone https://github.com/Scriddie/Varsortability
git clone https://github.com/kevinsbello/dagma
git clone https://github.com/ignavierng/golem
git clone https://github.com/fishmoon1234/DAG-NoCurl NoCurl

Make sure to have R installed and that the path to the R directory in file experiment.py is correct. Also note the required packages on R for the causal discovery toolbox.

Execution

# run all experiments
bash run.sh
# plots and tables
bash report.sh