Differentiable Causal Discovery from Interventional Data

Repository for "Differentiable Causal Discovery from Interventional Data".

• The code for DCDI is available here.

• The datasets used in the paper are available here

The code for the following baselines is also provided:

• CAM
• GIES
• IGSP and UT-IGSP

To run DCDI, you can use a command similar to this one: python ./dcdi/main.py --train --data-path ./data/perfect/data_p10_e10_n10000_linear_struct --num-vars 10 --i-dataset 1 --exp-path exp --model DCDI-DSF --intervention --intervention-type perfect --intervention-knowledge known --reg-coeff 0.5

Here, we assume the zip files in the directory data have been unzipped and that the results will be saved in the directory exp. With this command, you will train the model DCDI-DSF in the perfect known setting (with reasonable default hyperparameters) using the first instance (i_dataset = 1) of the 10-node graph dataset with linear mechanisms and perfect intervention. For further details on other hyperparameters (e.g. the architecture of the networks), see the main.py file where hyperparameters have a description.