source: arxiv machine learning: a survey on federated causal discovery and inference
level: research
causal reasoning helps with data-driven decisions by finding causal structures and estimating effects. often, data sits across different organizations and cannot be combined due to privacy rules or communication limits. federated learning lets parties work together without sharing raw data. this has led to new work in federated causal discovery and federated causal inference. the field is growing fast but lacks a clear overview, making it hard for new researchers to start.
the paper organizes existing methods around three main design choices. first, how causal structures are learned, such as constraint-based, score-based, or functional causal model approaches. second, how data is split across parties, like horizontal, vertical, or mixed partitions. third, what structural knowledge each party gets, from full graph access to limited local views. the survey also looks at privacy protections, communication costs, and how methods handle different data types.
federated causal inference is covered separately, focusing on estimating treatment effects without centralizing data. the review highlights gaps like handling non-iid data, missing values, and real-world deployment. it also points out the need for better benchmarks and standardized evaluation. the authors aim to give a clear map of the field and suggest future directions for both theory and practice.
why it matters: it helps ai and data science practitioners understand how to learn causal relationships from distributed data without violating privacy, which is key for healthcare, finance, and other sensitive domains.
source: arxiv machine learning: a survey on federated causal discovery and inference