Abstract
Federated Causal Discovery (FCD) enables institutions to collaboratively learn causal models without sharing raw data. A critical challenge in FCD is sample quality heterogeneity, where local data quality varies across variable sets. We introduce FedACD (Federated Adaptive Causal Discovery), which utilizes an adaptive masking mechanism to filter unreliable local findings. By ignoring low-quality local results, FedACD prevents global model degradation. Our evaluation shows that FedACD improves F1 scores by 19.65% and reduces Structural Hamming Distance (SHD) by 23.57% on average compared to state-of-the-art methods.
Introduction
Causal discovery identifies cause-and-effect relationships from observational data, a task traditionally requiring centralized data access. In privacy-sensitive domains like healthcare, data is often siloed across institutions that cannot share raw information due to legal and ethical constraints. Federated Causal Discovery (FCD) addresses this by allowing clients to learn a global model collaboratively.
However, existing FCD methods are vulnerable to sample quality heterogeneity: a client might provide high-quality data for some variables but noisy or "bad" data for others. Aggregating these unreliable local insights poisons the global model, degrading overall accuracy. FedACD addresses this by identifying and masking unreliable local results without compromising privacy, ensuring that only high-quality information contributes to the global consensus.
Methodology: The FedACD Framework
FedACD operates in two primary phases to construct the global causal graph while maintaining data privacy:
- Federated Causal Skeleton Learning (FCSL): Clients perform local Conditional Independence (CI) tests to find potential connections between variables. The core innovation is Adaptive Masking: clients calculate p-values for their CI tests and mask the most unreliable edges (those with low p-values, indicating potential errors) with a value of "-1". The server then aggregates only the unmasked "votes," ensuring that the global skeleton is built from the most reliable local information.
- Federated Skeleton Orientation (FSO): Once the undirected skeleton is established, clients use score-and-search methods (e.g., BDeu for discrete data or BIC for continuous data) to determine edge directions. The server aggregates these local directed acyclic graphs (DAGs) to finalize the global model.
Experimental Evaluation
We evaluated FedACD against five state-of-the-art baselines (e.g., FedPC, NOTEARS-ADMM) using benchmark Bayesian networks (Child, Insurance, Alarm), synthetic non-IID datasets, and real-world lung cancer gene expression data.
Performance Metrics
The framework was assessed using the F1 score (measuring precision and recall) and Structural Hamming Distance (SHD), which counts the edits needed to reach the true graph. FedACD consistently outperformed all baselines across various scenarios. For instance, on the Child dataset with 20 clients, FedACD achieved an F1 score 42% higher than the best-performing baseline.
Key Insights and Conclusion
The success of FedACD stems from the "less is more" principle: by allowing clients to withhold uncertain information via adaptive masking, the global model relies on the highest-quality local data available for each specific variable pair. This strategy is particularly effective as the number of clients increases and local data becomes sparser, where standard methods typically collapse.
While FedACD represents a significant advancement, it currently assumes causal sufficiency (no hidden variables) and focuses on horizontal federated learning. Future work will explore more complex scenarios involving latent variables and the integration of interventional data.
Summary updated