Handling Covariate Mismatch in Collaborative Linear Prediction
The research develops a framework for federated learning when clients see different data subsets (covariate mismatch). It proposes two main solutions: plug-in methods for low dimensions and an Impute-then-Regress strategy for high dimensions. The findings show that in high dimensions, Impute-then-Regress beats local learning, and in low dimensions, the plug-in estimator is preferred.