Variational Causal Discovery Unconstrained by Acyclicity (VCUDA)

This is the implementation for our paper: Scalable Variational Causal Discovery Unconstrained by Acyclicity, accepted at ECAI 2024.

Setup • Structure • Experiments • Citation

Setup

conda env create -n vcuda --file env.yml
conda activate vcuda

You can use env-cpu.yml if you don't have a gpu.

Structure

• castle/: the modified gcastle package.
• data/: synthetic and real datasets. Synthetic datasets will be generated when the experiment runs.
• exp/: code for running the experiments.
• src/: source code for VCUDA and baselines.

Experiments

Synthetic dataset notation

We denote a synthetic dataset as {ER/SF}_d{nodes}_e{edges}_N{samples}_{noisetype/nonlinear functions}.

For example:

• ER_d10_e10_N1000_gauss: generated from the ER graph model of 10 nodes and 10 expected edges (degree of 1) with 1000 samples from the linear model with Gaussian noises.
• SF_d10_e10_N1000_gp: generated from the SF graph model of 10 nodes and 10 expected edges (degree of 1) with 1000 samples from the Gaussian Process.

DAG sampling

The results reported in the paper are saved in times.csv.

To generate Figure 2 in the paper:

python exp/dag_sampling/timing.py

The plot is saved in exp/dag_sampling/plot.pdf.

Add --run option to reproduce the result.

To reproduce the results of differentiable DAG sampling optimization:

python exp/dag_sampling/main.py --run

DAG learning

The results reported in the paper are saved in folder exp/dag_learning/results/.

To reproduce the results, uncomment the run: false for a method that you want to rerun in the config files in exp/dag_learning/configs/ and run:

python exp/dag_learning/run.py --cfg_file linear.yml

Use nonlinear.yml and real.yml to produce the results on nonlinear and real datasets, respectively.

To generate Figure 3, 4:

python exp/dag_learning/viz.py --cfg_file linear.yml 

The resultant plots are saved in exp/dag_learning/plot/ following the format linear/nonlinear-{d nodes}.pdf

In case you got an error of "Tensors in different devices", refer to this issue.

Citation

@inproceedings{hoang2024enabling,
      title={Scalable Variational Causal Discovery Unconstrained by Acyclicity}, 
      author={Hoang, Nu and Duong, Bao and Nguyen, Thin},
      year={2024},
      booktitle = {European Conference on Aritificial Intelligence (ECAI)},
}