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

Fed-CausalDiff is a federated causal diffusion framework designed for do-simulation and policy evaluation, addressing the limitations of traditional federated learning approaches that rely solely on historical data. The architecture features a global causal score function and a local confounding score function, enabling decoupled synchronization (DSS) to manage heterogeneity across clients. Experimental results on four datasets show improved average treatment effect (ATE) and policy-value estimation accuracy, highlighting its potential for enhancing inference fidelity while reducing communication costs in federated settings.