CausalDisco 🪩

CausalDisco provides baseline algorithms and analytics tools for Causal Discovery in Python. The package can be installed by running pip install CausalDisco. Additional information can be found in the documentation.

Baseline Algorithms

Find the following baseline algorithms in CausalDisco/baselines.py

• R²-SortnRegress
• Var-SortnRegress
• Random-SortnRegress

Analytics Tools

Find the following analytics tools in CausalDisco/analytics.py

• R²-sortability
• Var-sortability
• order_alignment

Sources

If you find our algorithms useful please consider citing

• Beware of the Simulated DAG!
• A Scale-Invariant Sorting Criterion to Find a Causal Order in Additive Noise Models.

@inproceedings{reisach2023scale,
    title = {{A Scale-Invariant Sorting Criterion to Find a Causal Order in Additive Noise Models}},
    author = {Alexander G. Reisach and Myriam Tami and Christof Seiler and Antoine Chambaz and Sebastian Weichwald},
    booktitle = {{Advances in Neural Information Processing Systems 36 (NeurIPS)}},
    year = {2023},
    doi = {10.48550/arXiv.2303.18211},
}

@inproceedings{reisach2021beware,
    title = {{Beware of the Simulated DAG! Causal Discovery Benchmarks May Be Easy to Game}},
    author = {Alexander G. Reisach and Christof Seiler and Sebastian Weichwald},
    booktitle = {{Advances in Neural Information Processing Systems 34 (NeurIPS)}},
    year = {2021},
    doi = {10.48550/arXiv.2102.13647},
}

A Simple Example

# --- sample data from a linear SCM:

import numpy as np
from scipy import linalg

d = 10
W = np.diag(np.ones(d-1), 1)
X = np.random.randn(10000, d).dot(linalg.inv(np.eye(d) - W))
X_std = (X - np.mean(X, axis=0))/np.std(X, axis=0)

# --- run analytics and print results:

from CausalDisco.analytics import (
    var_sortability,
    r2_sortability,
    snr_sortability
)

print(
    f'True\n{W}\n'
    f'var-sortability={var_sortability(X, W):.2f}\n'
    f'R^2-sortability={r2_sortability(X, W):.2f}\n'
    f'SNR-sortability={snr_sortability(X, W):.2f}'
)

# --- run baselines and print results:

from CausalDisco.baselines import (
    random_sort_regress,
    var_sort_regress,
    r2_sort_regress
)

print(
    f'True\n{W}\n'
    '--- randomRegress ---\n'
    f'Recovered:\n{1.0*(random_sort_regress(X)!=0)}\n'
    f'Recovered standardized:\n{1.0*(random_sort_regress(X_std)!=0)}\n'
    '--- varSortnRegress ---\n'
    f'Recovered:\n{1.0*(var_sort_regress(X)!=0)}\n'
    f'Recovered standardized:\n{1.0*(var_sort_regress(X_std)!=0)}\n'
    '--- r2SortnRegress ---\n'
    f'Recovered:\n{1.0*(r2_sort_regress(X)!=0)}\n'
    f'Recovered standardized:\n{1.0*(r2_sort_regress(X_std)!=0)}\n'
)

Shout out to Anne Helby Petersen for coming up with the "causal disco" naming idea first (check out her causalDisco package providing tools for causal discovery on observational data in R).