FedHF-Impute: Federated Imputation for Heterogeneous Feature Spaces

The newly developed federated imputation framework, FedHF-Impute, tackles the issue of varying feature spaces in tabular datasets. Conventional federated learning techniques, such as FedAvg, rely on aligned feature schemas, which become ineffective when clients have only partially overlapping feature sets. FedHF-Impute distinguishes between structural feature absence and typical missingness, utilizing a shared global feature graph to facilitate information exchange among statistically related features through message passing. This allows for indirect knowledge transfer across clients, even when features are not jointly observed, while maintaining standard federated communication. The framework is tailored for situations where parameter-averaging approaches yield minimal information transfer across weakly overlapping or entirely separate feature groups. Its effectiveness is demonstrated under simulated conditions of partial schema mismatch.