Fed-CausalDiff: Decoupled Synchronization for Federated Do-Simulation and Policy Evaluation

Federated learning (FL) is typically used for observational modeling, which is insufficient for causal inference and policy evaluation where sequential actions change future states. Standard FL methods often struggle with the time-varying confounding and statistical heterogeneity inherent in decentralized longitudinal data. The authors propose Fed-CausalDiff, a framework designed to perform interventional simulation (do-simulation) in a federated setting.

The core innovation is a structural decomposition of the latent state into a globally shared causal subspace and a locally retained confounding subspace. By using conditional score-based diffusion, the model learns to simulate how interventions affect future states. This architecture enables Decoupled Synchronization (DSS), where clients only share the parameters governing the causal dynamics with the server, while keeping site-specific confounding parameters local to handle non-IID data.

The authors evaluated Fed-CausalDiff on four datasets, including synthetic student interaction data and real-world clinical and bandit logs. The model demonstrated superior accuracy in estimating Average Treatment Effects (ATE) and policy values compared to existing federated generative and counterfactual baselines. By isolating shared causal laws from local nuisance factors, the framework achieves a more favorable trade-off between communication efficiency and the fidelity of counterfactual rollouts, making it robust to the statistical disparities commonly found in multi-institutional data.

This work bridges the gap between federated learning and causal inference. In sensitive domains like healthcare or financial risk management, data cannot be centralized due to privacy regulations. Fed-CausalDiff allows organizations to collaboratively learn the causal effects of policies without sharing raw data or individual-level records. This enables more reliable offline policy evaluation and decision-making in decentralized environments where traditional correlation-based models would otherwise introduce significant bias.