Advances in missing data imputation over federated data networks

Real-world tabular data are often fragmented across organizations and pervasively containmissing values. In distributed settings, centralized imputation is often infeasible due to privacy, regulatory, and institutional constraints. This dissertation studies missing-data imputation in federated data networks and develops method-level, system-level, and fairness-oriented solutions for both horizontally and vertically partitioned data.For horizontal federated settings, we propose two ICE-based methods for complex and heterogeneous missingness. First, Cafe-ICE addresses MNAR-dominant heterogeneous federations by modeling cross-client complementarity in missingness patterns and using complementarity-adjusted aggregation to improve client-level imputation. Second, KT-ICE addresses mechanism-type heterogeneity by transferring inter-feature knowledge from simple-missingness (MCAR) clients to non-MCAR clients through a propensity-aware personalized combination strategy. For vertical federated settings, where features are split across parties for aligned entities, we develop vSGDICE, an SGD-based ICE framework that supports mixed-type data and exchanges only aggregated intermediate information rather than raw features.To improve reproducibility and benchmarking in this emerging area, we develop FedImpute, an open-source toolkit that standardizes scenario construction, federated imputation execution, and evaluation. Beyond imputation performance, the dissertation also examines fairness implications through a controlled, centralized study, showing that fairness impacts depend not only on the missing rate but also on the missingness structure with respect to sensitive attributes and outcomes.Extensive experiments on synthetic, semi-simulated, and real-world datasets show that the proposed methods consistently improve imputation quality and downstream utility relative to existing baselines while preserving federated constraints. Together, these contributions provide a practical foundation for robust, privacy-conscious, and fairness-aware handling of missing data in distributed environments such as healthcare and finance.