Regret-Based Federated Causal Discovery with Unknown Interventions

Researchers introduce I-PERI, a federated causal discovery algorithm that handles unknown client-level interventions across decentralized systems. The method advances privacy-preserving causal inference by recovering tighter equivalence classes when clients operate under heterogeneous, undisclosed policies—addressing a critical gap between theoretical causal discovery methods and real-world deployment constraints.

Causal discovery—the process of inferring cause-and-effect relationships from data—has traditionally assumed homogeneous environments with known experimental conditions. This research tackles a fundamental limitation of existing federated causal discovery methods: the unrealistic assumption that all participating clients share identical causal models. In healthcare networks, financial institutions, or sensor networks, each participant operates under distinct protocols and policies that introduce unknown interventions, making standard approaches invalid.

I-PERI represents a methodological advance by explicitly modeling heterogeneous interventions while preserving privacy across federated clients. The algorithm first identifies the union of causal structures across all clients, then leverages structural differences induced by varying interventions to orient additional edges in the causal graph. This produces a Φ-Markov Equivalence Class—a narrower set of possible causal structures than traditional methods yield.

For practitioners in healthcare, finance, and distributed systems, this development matters substantially. Federated settings increasingly dominate real-world applications due to data privacy regulations and organizational silos. Current methods either ignore intervention heterogeneity or require explicit disclosure of client-specific interventions, both problematic in practice. I-PERI's theoretical guarantees on convergence and privacy preservation suggest practitioners can recover more precise causal relationships without compromising institutional confidentiality.

The research opens possibilities for more robust causal inference in decentralized environments. Future work likely focuses on scaling these methods to high-dimensional settings and validating performance across diverse real-world domains where heterogeneous interventions are endemic.

• →I-PERI enables causal discovery in federated systems where clients operate under unknown, heterogeneous interventions—a realistic constraint previous methods ignored.
• →The algorithm recovers a tighter Φ-Markov Equivalence Class by exploiting structural differences induced by varying client-level policies.
• →Theoretical guarantees ensure both convergence properties and privacy preservation, making it suitable for sensitive domains like healthcare and finance.
• →This addresses a critical gap between idealized causal discovery assumptions and decentralized real-world deployments with institutional heterogeneity.
• →The method applies to any federated setting where participants maintain distinct protocols while collaborating on causal inference tasks.