gRESIT

This repo aims at learning and representing causal graphs based on grouped data. Theoretical details are presented in the paper

@misc{goebler2025,
  title={Nonlinear Causal Discovery for Grouped Data},
  author={Konstantin G\"obler and Tobias Windisch and Mathias Drton},
  year={2025},
  eprint={2506.05120},
  archivePrefix={arXiv},
  primaryClass={stat.ML},
  url={<https://arxiv.org/abs/2506.05120}>,
}

Authors

• Konstantin Göbler (TU Munich, Bosch)
• Tobias Windisch (HS Kempten)

Maintainer: Martin Roth (Bosch)

Table of contents

• Documentation
• How to install
• How to build
• How to use
• How to test
• Github Actions

Documentation

The documentation can be found here.

How to install

The package can be installed with

pip install gresit

How to build

Using the Makefile the package can be installed in an editable way like this:

make sync-venv

To use the pre-commit hooks, one has to enable them in the venv, by

pre-commit install

Then these hooks are excecuted before every commit. You can run the hooks for all files also separately

pre-commit run --all-files

or to disable the pip-compile hook, which takes some time

SKIP=pip-compile pre-commit run --all-files

or equivalent

make pre-commit

How to use

Consider the following example. We refer to the documentation for more detailed information.

Generating Synthetic Data

We first generate synthetic data using an Erdős–Rényi random graph model. Each group of variables is defined with a specified size and edge density.

from gresit.synthetic_data import GenERData

data_gen = GenERData(
    number_of_nodes=10,
    group_size=2,
    edge_density=0.2,
)

data_dict, _ = data_gen.generate_data(num_samples=1000)

The output data_dict is a dictionary where each key corresponds to a group, and the values are the observed samples.

Fitting a Graph Model

We now fit a gRESIT model using Multioutcome_MLP as the regressor and HSIC as independence test.

from gresit.group_resit import GroupResit
from gresit.independence_tests import HSIC
from gresit.torch_models import Multioutcome_MLP

model = GroupResit(
    regressor=Multioutcome_MLP(),
    test=HSIC,
    pruning_method="murgs",
)
learned_dag = model.learn_graph(data_dict=data_dict)

# Show the learned graph:
learned_dag.show()
# or show interactive mode:
model.show_interactive()

Accessing the Learned Graph

The learned adjacency matrix representing the estimated group-level graph and a causal ordering can be accessed via:

model.adjacency_matrix
model.causal_ordering

How to test

In general we use pytest and the test suite can be executed locally via

python -m pytest

Github Actions

Documentation with mkdocs

We use mkdocs for building the documentation, this is the corresponding workflow.

Automated issue workflow

With this workflow newly created issues are automatically added to our MFD2 project.

Pre-commit

With this workflow the pre-commit rules, specified in .pre-commit-config.yaml, are executed.

To use pre-commit locally, please use

pre-commit install

Testing

With this workflow the tests are executed.

Third-Party Licenses

Runtime dependencies

Name	License	Type
numpy	BSD-3-Clause License	Dependency
pandas	BSD-3-Clause License	Dependency
scikit-learn	BSD-3-Clause License	Dependency
statsmodels	BSD-3-Clause License	Dependency
plotly	MIT License	Dependency
xgboost	Apache License 2.0	Dependency
torch	BSD-3-Clause License	Dependency
seaborn	BSD-3-Clause License	Dependency
pyspark	Apache License 2.0	Dependency
scikit-misc	BSD-3-Clause License	Dependency
gadjid	MIT License	Dependency
tqdm	MIT License	Dependency
dcor	MIT License	Dependency
llvmlite	BSD-2-Clause License	Dependency
causal-learn	MIT License	Dependency
gcastle	Apache License 2.0	Dependency
gpytorch	MIT License	Dependency

Development dependency

Name	License	Type
mike	BSD-3-Clause License	Optional
mkdocs	BSD-2-Clause License	Optional
mkdocs-material	MIT License	Optional
mkdocstrings	ISC License	Optional
pip-licenses	MIT License	Optional
pip-tools	BSD-3-Clause License	Optional
pre-commit	MIT License	Optional
pytest	MIT License	Optional
pytest-cov	MIT License	Optional
ruff	MIT License	Optional
uv	MIT License	Optional