This is the implementation of the paper Differentiable Invariant Causal Discovery
- torch == 1.9.0+cu111
- Numpy
- python3
We have put the example data of 10 nodes and 40 edges under linear setting in the folder data. The full dataset could be downloaded here.
You can run NOTEARS and DICD for the linear experiments for ER4 graph with 10 nodes via the following codes:
python linear_exp.py --s0 40 --d 10 --method NOTEARS --graph_type ER
python linear_exp.py --s0 40 --d 10 --method DICD --graph_type ER
You can run NOTEARS and DICD for the nonlinear experiments for SF4 graph with 10 nodes via the following codes:
python nonlinear_exp.py --s0 40 --d 10 --method NOTEARS --graph_type SF
python nonlinear_exp.py --s0 40 --d 10 --method DICD --graph_type SF
For DAG-GNN and DAG-NoCurl, We follow the official implementations with the link provided as follows:
- DAG-GNN: https://github.com/fishmoon1234/DAG-GNN
- DAG-NoCurl: https://github.com/fishmoon1234/DAG-NoCurl
the scripts are as follows:
python linear_exp.py --s0 40 --d 10 --method DAG-GNN --graph_type ER
python linear_exp.py --s0 40 --d 10 --method NoCurl --graph_type ER
python linear_exp.py --s0 40 --d 10 --method DARING --graph_type ER
python non_linear_exp.py --s0 40 --d 10 --method DAG-GNN --graph_type ER
python non_linear_exp.py --s0 40 --d 10 --method NoCurl --graph_type ER
python non_linear_exp.py --s0 40 --d 10 --method DARING --graph_type ER