Federated Causal Discovery Under Unknown Client Interventions
The I-PERI algorithm, a novel federated causal discovery approach, effectively identifies a more precise equivalence class from decentralized datasets that involve unknown client-level interventions. Traditional methods often presume uniform causal models among clients, which is not practical, especially in varied settings like hospitals with distinct protocols. Initially, I-PERI reconstructs the CPDAG from the collective client graphs and subsequently directs additional edges by leveraging the structural variations caused by interventions. This process results in the Φ-Markov Equivalence Class, offering a more sophisticated representation compared to the conventional Markov equivalence class. This research tackles issues related to data decentralization and privacy without needing prior knowledge of intervention targets.
Key facts
- I-PERI is a federated algorithm for causal discovery.
- It handles unknown client-level interventions.
- It recovers the CPDAG of the union of client graphs.
- It orients additional edges using structural differences from interventions.
- The resulting equivalence class is called the Φ-Markov Equivalence Class.
- The method addresses data decentralization and privacy constraints.
- Existing methods assume all clients share the same causal model.
- The work is motivated by heterogeneous protocols across hospitals.
Entities
—