Federated causal structure learning with missing data

Abstract

Federated causal structure learning (CSL) is an emerging research direction that aims to discover causal relationships from decentralized data across multiple clients, while preserving data privacy. Existing federated CSL algorithms primarily focus on complete datasets and often overlook data-quality issues, such as missing data, which are common in real-world scenarios. Moreover, client diversity can destabilize federated CSL, and this challenge is further worsened by missing data. To address these issues, we propose FedImpCSL, a novel federated CSL method, for effectively handling missing data. Our approach consists of two key components: (1) a local-to-global missing data imputation strategy that reconstructs imputed and accurate datasets from missing samples, and (2) a dynamic client weighting and weighted aggregation strategy to address inter-client differences, enhancing the CSL accuracy without utilizing each client’s original data. We demonstrate the effectiveness of FedImpCSL through comprehensive experiments on various types of datasets, showing its superior performance over existing federated CSL methods in handling missing data scenarios.