This repository accompanies the paper Effects of Distributional Biases on Gradient-Based Causal Discovery in the Bivariate Categorical Case.

This repository was developed on Python 3.10. The required packages are:

• torch
• tqdm
• numpy
• matplotlib
• pathlib
• gitpython
• scipy
• pandas

When creating videos also:

• ffmpeg

You can configure model parameters in utils/model.py. With run.py you can run a model, which will automatically create a folder plots/ if it does not exist and then a folder for the current experiment inside it. It will save various plots and data there.

For performing multiple runs with different parameters, for instance different epsilon, you can use generate_run_data.py for this.

If you want to investigate distributions, use the notebooks in notebooks/.