Federated Imputation under Heterogeneous Feature Spaces

The paper's core contribution is separating the problem of missing features (columns a client doesn't hold) from missing values (sparse entries within held columns). Prior federated learning work conflated these, forcing all clients to share identical schemas. FedHF-Impute uses a learned feature graph to infer correlations across organizational boundaries, enabling imputation without exposing raw records.

This connects directly to the attention dispersion work from mid-May on dynamic graph transformers. Both papers identify failure modes that emerge when real-world data violates standard assumptions (temporal distribution shifts there, heterogeneous schemas here) and propose structural fixes rather than post-hoc patches. The feature graph approach here mirrors the conditional-marginal discretization logic in the flow sampler paper (same release date), where decoupling independent concerns (bridge geometry vs. marginal dynamics there, feature presence vs. value sparsity here) yields cleaner solutions. For practitioners, this matters because enterprise data rarely fits the homogeneous-schema assumption that most federated learning libraries encode.

If FedHF-Impute ships with open-source code and someone demonstrates it on a real multi-organization financial dataset (loan applications across lenders with different underwriting columns) within six months, that signals the technique is production-ready. If adoption stalls at academic benchmarks without industry pilots, the feature correlation assumptions may not hold on messy real data.