Federated Causal Discovery

Adapts classical constraint-based causal discovery methods, such as the PC algorithm and FCI (Fast Causal Inference), to operate across distributed data partitions. Instead of pooling raw data to perform conditional independence tests, each client computes local sufficient statistics—such as covariance matrices or frequency counts—which are then securely aggregated. The global server reconstructs the skeleton graph and orients edges using the aggregated statistical summaries, ensuring that individual-level data never leaves its origin site. This approach is particularly effective for cross-silo settings where each institution holds a statistically meaningful sample size.

Implements score-based causal discovery algorithms like Greedy Equivalence Search (GES) or NOTEARS in a federated manner. Each client computes a local score for candidate causal graphs—typically a BIC score or BDeu score—based on its private data. These local scores are transmitted to a central coordinator that sums them to obtain the global score function. The optimization then proceeds to search over the space of Directed Acyclic Graphs (DAGs) using the aggregated score, finding the structure that best explains the combined data. This method naturally handles non-IID data distributions across sites, as score decomposability allows for principled aggregation.

Integrates differential privacy (DP) mechanisms directly into the federated causal discovery pipeline to provide formal privacy guarantees. Before transmitting local statistics or scores, each client injects calibrated Gaussian or Laplacian noise scaled to a predefined privacy budget (ε, δ). This ensures that the output causal graph does not leak information about any single individual's record. The challenge lies in balancing the privacy-utility tradeoff: excessive noise can mask weak causal signals or introduce spurious edges, while insufficient noise fails to provide meaningful protection. Advanced techniques include subsampling and moment accounting to tighten privacy loss bounds.

Addresses the critical challenge of causal heterogeneity across federated sites, where the true underlying causal mechanism may differ between populations. Techniques include:

• Federated causal clustering: Grouping clients with similar causal structures before performing joint discovery
• Meta-learning for causal discovery: Learning a shared prior over causal structures that can be rapidly adapted to site-specific data
• Invariant causal prediction across environments: Identifying causal parents that consistently predict an outcome across all client sites, leveraging distributional shifts as a signal rather than a nuisance This capability is essential for multi-center clinical studies where treatment effects may vary across demographic groups.

Employs secure multi-party computation (SMPC) protocols to aggregate the sufficient statistics required for causal discovery without revealing any individual client's contribution. Using techniques such as secret sharing and secure aggregation via pairwise masking, the central server can compute the sum of local covariance matrices or contingency tables while remaining cryptographically blinded to individual inputs. This provides stronger protection than simple federated averaging, as it defends against honest-but-curious servers that might attempt to reconstruct private data from aggregated statistics. The computational overhead is manageable for the relatively low-dimensional sufficient statistics used in causal discovery.

Extends beyond constraint-based and score-based methods to functional causal models (FCMs) such as LiNGAM (Linear Non-Gaussian Acyclic Model) and additive noise models. These methods exploit distributional asymmetries to identify the full causal direction—not just the Markov equivalence class—without requiring experimental interventions. In the federated setting, clients compute local pairwise measures of independence between residuals and potential causes, which are then aggregated to determine edge directions. This approach is particularly valuable for biomarker discovery, where distinguishing cause from mere correlation is critical for identifying therapeutic targets.

A fundamental challenge in federated causal discovery where local datasets are statistically heterogeneous. A variable that acts as a confounder in one hospital's population may be a mediator in another's. This heterogeneity can violate the causal Markov condition assumed by many algorithms, requiring federated methods that explicitly model distribution shift rather than naively pooling conditional independence test results.

A prerequisite for valid multi-site causal inference. Systematic technical variation across sites can create spurious causal edges that reflect laboratory protocols rather than biology. Federated batch correction harmonizes data distributions without centralization, ensuring that discovered causal relationships are driven by genuine biological mechanisms rather than site-specific artifacts.