Federated Causal Discovery from Interventions
Submission to the 4th Conference on Causal Learning and Reasoning
Submission to the 4th Conference on Causal Learning and Reasoning
Anonymous source code: please refer to the README file for reproducibility instructions.
You can unpack the data following the instructions in the main repository.
Existing causal discovery methods typically require the data to be available in a centralized location. However, many practical domains, such as healthcare, limit access to the data gathered by local entities, primarily for privacy and regulatory constraints. We propose FedCDI, a federated framework for discovering causal structures from distributed datasets containing observational and interventional data.
Clients' contributions are rated by the extent of their access to interventional data and the location of interventions in the overall causal structure, owing to our novel knowledge aggregation method. Therefore, a client presents a stronger vote on the existence of a cause-effect relationship near its intervened covariates.
Figure 1. Overview of our federated setup and the role of the proximity-based aggregation method.
Figure 2.1 FedCDI can compete with a centralized approach while preserving clients' privacy.
Figure 2.2 The proximity-based aggregation outperforms any naive voting-based method previously developed.
Figure 4. The first two rows are clients trying to recover a chain graph. The third row is the federated aggregation outcome. Note how consensus is reached through privacy-preserving collaboration among clients. The consensus is evident from prediction entropy analysis in the second figure as well.