Baseline

Here, we describe the baseline models and evaluation metrics used.

Models

Examples to run baselines can be found in notebooks/eval_pipeline.ipynb. Follow installation and (more) runtime instructions of each baseline in the provided github links.

• PCMCI+: jakobrunge/tigramite

• FPCMCI: lcastri/fpcmci

• VARLiNGAM: cdt15/lingam

• RCD: py-why/causal-learn

• GIN: py-why/causal-learn

• GRaSP: py-why/causal-learn

• DYNOTEARS: mckinsey/causalnex

• Neural GC: iancovert/Neural-GC

• CUTS+: jarrycyx/unn

• TSCI: KurtButler/tangentspace

• TCDF: M-Nauta/TCDF

Metrics

We use AUROC and AUPRC scores to evaluate the accuracy of discoverd summary graph. The structural Hamming distance (SHD) of the thresholded graph is reported as well.

Installation

The baselines need optional dependencies. Install them with:

pip install "causaldynamics[baselines]"
# FPCMCI and PCMCI+ also need IDTxl, which is only available from GitHub
pip install cython
pip install --no-build-isolation git+https://github.com/pwollstadt/IDTxl.git

DYNOTEARS (causalnex) requires pandas<2 and an outdated numpy pin, so install it in a separate environment:

pip install causaldynamics "pandas<2" "pgmpy<0.1.20" pathos pyvis
pip install --no-deps "causalnex @ git+https://github.com/mckinsey/causalnex.git@develop"

From a clone of the repository, use uv sync --extra baselines --group idtxl, or uv sync --group causalnex for DYNOTEARS. These install the exact versions from uv.lock.