Federated Learning of Generalized Linear Causal Networks

Area of Science:

• Causal inference and machine learning.
• Development of privacy-preserving AI algorithms.

Background:

• Causal discovery is crucial for scientific understanding but faces challenges with distributed data.
• Data privacy concerns necessitate decentralized approaches to data analysis.

Purpose of the Study:

• To propose a novel federated causal discovery method, DARLS (distributed annealing on regularized likelihood score).
• To enable learning causal graphs from decentralized data while ensuring privacy.

Main Methods:

• DARLS employs a distributed annealing process to explore topological sorts.
• It uses distributed optimization with multi-round communication for graph structure estimation.
• New identifiability results for causal graphs with generalized linear models are derived.

Main Results:

• DARLS converges to the optimal solution achievable with pooled data (an oracle).
• It is the first distributed method offering finite-sample oracle guarantees for causal graph learning.
• Simulations and real-world applications show DARLS outperforms existing federated methods.

Conclusions:

• DARLS effectively addresses causal discovery in distributed, privacy-sensitive settings.
• The method demonstrates significant advantages over current federated learning approaches.
• DARLS provides a robust solution for estimating causal networks from decentralized data.