This is the implementation of our paper Supervised Full DAG Causal Discovery. If this is helpful for your research, please consider cite our paper.

The experiments are sketched in src/main.jl. To run it, you first instantiate the project invironment, then run

julia --project src/main.jl

Or you can use the pre-trained models. Pre-trained models can be found at:

• d=10 pre-trained model
• d=20 pre-trained model
• d=50 pre-trained model
• d=100 pre-trained model

You can also found them in the release page.

Load the pretrained models:

@load "/path/to/model.bson" model

Data can be generated by:

spec = DataSpec(d=10, k=1, gtype=gtype, noise=:Gaussian)
ds, test_ds = spec2ds(spec)

Apply the model on the data:

x, y = next_batch!(test_ds)
pred = model(x)
loss = myσxent(pred, y)
metric = sup_graph_metrics(cpu(σ.(pred)), cpu(y))
@show metrics

The script for running other method is src/baseline.py. You should install those methods (various python and R projects) before running the script.