[ICLR 2024] Federated Causal Discovery from Heterogeneous Data

This is the official implementation of the paper "Federated Causal Discovery from Heterogeneous Data", ICLR 2024.

If you find it useful, please consider citing:

@inproceedings{li2024learning,
  title={Learning to Optimize Permutation Flow Shop Scheduling via Graph-based Imitation Learning},
  author={Li, Loka and Ng, Ignavier and Luo, Gongxu and Huang, Biwei and Chen, Guangyi and Liu, Tongliang and Gu, Bin and Zhang, Kun},
  booktitle={International Conference on Learning Representations},
  year={2024}
}

Overview

• In this paper, we propose FedCDH, a novel constraint-based approach for federated causal discovery from heterogeneous data. The figure below exhibits the overall framework of our FedCDH.

How to Run

• Installation: R package.

  • Our federated conditional independet test method is developed based on R Package, please follow their procedures and install all the dependencies at first.

• Installation: Environment.

# Set up a new conda environment with Python 3.8.
conda create -n FedCDH python=3.8
conda activate FedCDH

# Please navigate to the root directory, and install our source code.
pip install -e . 

# Install other python libraries.
pip install causaldag rpy2 numpy scipy tqdm networkx 

• Evaluation: quick start.

# Parameters:
#     N: number of instances to evaluate
#     d: number of variables
#     K: number of clients
#     n: number of samples in one client
#     model: data generation model, linear or general.
cd tests
python TestFedCDH.py --N 10 --d 6 --K 10 --n 100 --model linear  

Acknowledgements

We would like to sincerely thank these related works and open-sourced codes which inspired us:

• Causal-learn package: https://github.com/py-why/causal-learn.
• CausalDAG package: https://uhlerlab.github.io/causaldag.
• Huang et al., Causal Discovery from Heterogeneous/Nonstationary Data with Independent Changes. JMLR, 2020.
  • Code: https://github.com/Biwei-Huang/Causal-Discovery-from-Nonstationary-Heterogeneous-Data.
• Zhang et al., Kernel-based Conditional Independence Test and Application in Causal Discovery. UAI, 2011.
  • Code: https://causal-learn.readthedocs.io/en/latest/independence_tests_index/kci.html.
• Strobl et al., Approximate Kernel-Based Conditional Independence Tests for Fast Non-Parametric Causal Discovery. Journal of Causal Inference, 2018.
  • Code: https://github.com/ericstrobl/RCIT.

And we also sincerely thank the authors of the baseline methods for making their source codes public:

• Gao et al., FedDAG: Federated DAG Structure Learning. TMLR, 2022.
  • Code: https://github.com/ErdunGAO/FedDAG.
• Ng et al., Towards Federated Bayesian Network Structure Learning with Continuous Optimization. AISTATS, 2022.
  • Code: https://github.com/ignavierng/notears-admm.