CaPS

Code and datasets of paper "Ordering-Based Causal Discovery for Linear and Nonlinear Relations". (NeurIPS 2024)

Install dependencies

All dependencies can be installed with the following command. Note that the cdt library is dependent on the R environment and should be installed according to its documentation.

pip install -r ./requirements.txt

Examples of command

Here we give examples of command for both real and synthetic data. Please change the settings in the following for what you need.

An example for real data.

python3 train_order.py --dataset sachs

An example for synthetic data.

python3 train_order.py --dataset SynER1 --linear_rate 1.0

Settings

The important setting and its default value are given in the following table.

Name	Default	Description
dataset	sachs	It will automatically generate synthetic data if dataset starts with 'Syn', otherwise it reads the real data directly.
linear_rate	1.0	The linear proportion of synthetic data. 0.0 means all relations are nonlinear and 1.0 means all relations are linear.
linear_sem_type	gauss	The linear SEM of synthetic data, containing gauss, laplace and gumbel.
nonlinear_sem_type	gp	The nonlinear SEM of synthetic data, containing gp, mlp and mim.
pre_pruning	True	Use pre-pruning. Details are in Section 4.3 and Appendix B.
add_edge	True	Use edge supplement. Details are in Section 4.3 and Appendix B.
lambda1	50.0	The hyperparameter $\lambda$ for pre_pruning. Details are in Section 4.3 and Appendix B.
lambda2	50.0	The hyperparameter $\lambda$ for edge supplement. Details are in Section 4.3 and Appendix B.