Abstract
Federated causal discovery (FCD) aims to uncover causal relationships among variables from decentralized data across multiple clients, while preserving data privacy. In practice, the sample quality of each client's local data may vary across different variable spaces, referred to as sample quality heterogeneity. Thus, data from different clients might be suitable for learning different causal relationships among variables. Model aggregated under existing FCD methods requires the entire model parameters from each client, thereby being unable to handle the sample quality heterogeneity issue. In this paper, we propose the Federated Adaptive Causal Discovery (FedACD) method to bridge this gap. During federated model aggregation, it adaptively selects the causal relationships learned under the "good" variable space (i.e., one with high-quality samples) from each client, while masking those learned under the "bad" variable space (i.e., one with low-quality samples). This way, each client only needs to send the optimal learning results to the server, achieving accurate FCD. Extensive experiments on various types of datasets demonstrate significant advantages of FedACD over existing methods.
Introduction
Causal Discovery (CD) is fundamental for understanding complex systems by identifying directed relationships between variables, moving beyond simple correlation to find "what causes what." In privacy-sensitive domains like healthcare, data is often siloed across multiple institutions. Federated Causal Discovery (FCD) addresses this by enabling collaborative learning without raw data sharing.
However, real-world decentralized data often exhibits sample quality heterogeneity. For instance, one hospital might have high-quality equipment for measuring blood pressure but outdated tools for cholesterol. Existing FCD methods aggregate entire models from every client, allowing "bad" local data to pollute the global model. FedACD is the first method designed to address this by adaptively masking unreliable local results during the aggregation process.
Methodology: The FedACD Framework
The proposed FedACD framework introduces an adaptive variable space selection mechanism to mitigate sample quality heterogeneity. The framework operates in two distinct phases:
- Federated Causal Skeleton Learning (FCSL): Clients perform local Conditional Independence (CI) tests. The core innovation is Adaptive Masking: clients identify unreliable edges based on p-values. If a variable pair's data quality is low, the client masks that edge (setting it to -1) before sending the adjacency matrix to the server. The server then performs a voting aggregation, considering only unmasked, reliable edges.
- Federated Skeleton Orientation (FSO): With a global undirected skeleton established, clients use score-and-search methods (e.g., BDeu for discrete data or BIC for continuous data with Hill-climbing) to determine causal directions. The server aggregates these local Directed Acyclic Graphs (DAGs) to form the final model.
To ensure privacy, FedACD only exchanges structural information (adjacency matrices) and can be combined with additive homomorphic encryption to prevent the server from seeing individual client masks.
Experimental Results
FedACD was evaluated against five state-of-the-art baselines (including NOTEARS-ADMM and FedPC) across benchmark Bayesian networks (Child, Insurance, Alarm), synthetic non-IID datasets, and real-world gene expression data.
Key findings include:
- Accuracy Boost: FedACD improved the F1 score by an average of 19.65% over the best baselines.
- Error Reduction: It reduced the Structural Hamming Distance (SHD)—a metric for graph edit distance—by 23.57%.
- Robustness to Decentralization: On the Child dataset with 20 clients, FedACD’s F1 score was 42% higher than the best baseline, proving its effectiveness in highly decentralized, "data-starved" local scenarios where baselines typically fail due to false-positive edges.
Conclusion and Future Work
FedACD effectively addresses sample quality heterogeneity in federated causal discovery by allowing clients to selectively contribute reliable local insights. While the current model assumes causal sufficiency (no unobserved variables), future work will focus on handling latent confounders and integrating interventional data to further enhance discovery accuracy in complex real-world scenarios.
Summary updated