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

2026-06-21

AI Summary

Standard federated learning fails to support interventional reasoning and policy evaluation because it neglects the dynamic causal effects of actions on states. This work proposes Fed-CausalDiff, a federated causal diffusion framework that introduces a novel decoupled synchronization mechanism, decomposing latent state evolution into a global causal score function and local confounding score functions. This design reconciles cross-client causal consistency with local heterogeneity. By integrating causal inference, diffusion models, and federated learning within a dual-branch architecture, Fed-CausalDiff disentangles causal and confounding components, enabling do-calculus simulation and policy evaluation. Experiments demonstrate that Fed-CausalDiff significantly improves the accuracy of average treatment effect (ATE) and policy value estimation across four datasets, while achieving a favorable trade-off between communication overhead and inference fidelity.

Abstract

While federated learning enables collaborative modelling on decentralised data, standard methods merely fit historical observations. This purely observational approach is fundamentally insufficient for interventional inference and policy evaluation, as sequential actions dynamically alter future states. We propose \textbf{Fed-CausalDiff}, a federated causal diffusion framework for do-simulation. The architecture decomposes the evolution of the latent state into a global causal score function and a local confounding score function. This design enables \emph{decoupled synchronisation} (DSS), where clients aggregate only the shared causal mechanism while retaining site-specific confounders locally to handle heterogeneity. Experiments on four datasets demonstrate that Fed-CausalDiff achieves better ATE and policy-value estimation accuracy, offering a favorable trade-off between communication cost and inference fidelity.

Problem

• federated learning
• interventional inference
• policy evaluation
• causal inference
• do-simulation

Innovation

• federated causal inference
• diffusion models
• decoupled synchronization
• do-simulation
• policy evaluation