Towards Practical Federated Causal Structure Learning

Zhaoyu Wang, Pingchuan Ma, Shuai Wang

Abstract

Causal structure learning is essential for understanding complex systems, yet privacy constraints often prevent the centralization of sensitive data. This paper introduces FedC2SL, a robust federated framework for learning causal relationships from distributed data. By leveraging a novel federated conditional independence test and random projections, FedC2SL overcomes common pitfalls in federated causal discovery, such as restrictive noise assumptions, privacy leakage, and data heterogeneity. Experimental results demonstrate that FedC2SL significantly outperforms existing state-of-the-art methods in accuracy and scalability while providing strong privacy guarantees.

Introduction

Identifying causal "blueprints" is a fundamental task in scientific research. Traditionally, this requires a central repository of data, which is often unfeasible in privacy-sensitive domains like healthcare or finance. While Federated Learning (FL) allows model training on decentralized data, applying it to causal discovery is notoriously difficult. Existing methods frequently rely on unrealistic additive noise models, risk leaking sensitive local graph structures, and struggle with heterogeneous data distributions across clients.

Comparison of FedC2SL with existing federated causal discovery methods.

Methodology

FedC2SL (Federated Constraint-based Causal Structure Learning) shifts away from traditional score-based optimization toward a constraint-based approach. The framework's core innovation is a Federated Conditional Independence (CI) Test. Instead of aggregating complex local models or sharing biased local graphs, the server coordinates with clients to perform statistical tests (such as the $\chi^2$ test) globally.

To preserve privacy, FedC2SL utilizes random projections based on $\alpha$-stable distributions. This technique allows the server to estimate necessary global statistics—specifically the "second frequency moment"—without ever accessing raw data counts, ensuring client indistinguishability. The framework is versatile, offering two primary variants: FedPC for standard Directed Acyclic Graphs (DAGs) and FedFCI, which is designed to handle latent confounders (hidden variables) by representing the system as a Maximal Ancestral Graph (MAG).

A Maximal Ancestral Graph (MAG) where Z is a latent confounder influencing observed variables X and Y.

Evaluation and Results

The framework was rigorously evaluated using synthetic datasets, the Sachs protein signaling dataset, and the LUCAS dataset for downstream causal feature selection.

  • Superior Accuracy: On the Sachs dataset, FedC2SL achieved a Structural Hamming Distance (SHD) of 5.6, vastly outperforming the next best federated baseline, which lagged at 15.7.
  • Scalability: While competing methods like NOTEARS-ADMM saw performance collapse as variable counts reached 100, FedC2SL remained stable and accurate.
  • Robustness to Heterogeneity: FedC2SL is naturally resistant to client heterogeneity because it aggregates global statistics rather than averaging biased local views. In tests with diverse data distributions, it maintained accuracy where other methods saw error rates jump by an average of 4.7 points.
  • System Resilience: The protocol showed minimal performance degradation even when 20% of clients dropped out during the learning process.

Conclusion

FedC2SL provides a practical and private solution for federated causal discovery. By utilizing global CI tests and secure random projections, it offers a scalable, accurate path forward for causal inference in distributed environments without compromising sensitive information or relying on restrictive mathematical assumptions.

Summary updated

BibTeX
@inproceedings{wang2023towards,
  author = {Zhaoyu Wang and Pingchuan Ma and Shuai Wang},
  title = {{Towards Practical Federated Causal Structure Learning}},
  booktitle = {ECML/PKDD},
  year = {2023},
  doi = {10.1007/978-3-031-43415-0\_21},
}

Loading related content...

Figure