GoudetOlivier/CGNN

Replication code for the article "Learning Functional Causal Models with Generative Neural Networks"

Tensorflow Implementation of the CGNN

Code provided to reproduce the results from the article "Learning Functional Causal Models with Generative Neural Networks"

Requirements: numpy scipy scikit-learn tensorflow joblib pandas

In order to run the CGNN and launch the experiments:

1. First install the CGNN package. Enter in the code directory. Run the command line "python setup.py install develop --user"

2. Launch the example python script for pairwise inference: "python run_GNN_pairwise_inference.py"

3. Launch the example python script for graph reconstruction from a skeleton: "python run_CGNN_graph.py"

4. Launch the example python script for graph reconstruction in presence of hidden variables: "python run_CGNN_graph_hidden_variables.py"

5. The complete datasets used in the article may be found at the following url:

• pairwise datasets : http://dx.doi.org/10.7910/DVN/3757KX
• graph datasets : http://dx.doi.org/10.7910/DVN/UZMB69

Fast Pytorch implementation of CGNN available in the CDT

A faster implementation of CGNN in pytorch in available in the CausalDiscoveryToolBox (CDT)

https://github.com/Diviyan-Kalainathan/CausalDiscoveryToolbox

arXiv paper of the CDT: https://arxiv.org/abs/1903.02278