Hybrid federated learning improves reliability with missing data types
Hybrid Federated Learning for Multimodal Fusion under Missing Modalities and Site Heterogeneity
Machine Learning
Summary
When different sites have varied types of data that might sometimes be missing, it can be hard to combine everything to learn well. The authors developed a method that carefully balances the influence of different data types and accounts for missing pieces across locations. Their approach improves the worst-case performance at any site, which means the learning is more robust even when some data are not available or sites differ a lot. They tested this on simulations and real activity recognition data to show it helps especially under challenging conditions.
What this means in practice
- •For mobile app developers: Combine data from different sensors on users’ devices even when some sensors may be missing or behave differently across devices.
- •For healthcare analytics teams: Build models that integrate patient data from multiple sources while handling missing measurements and differences across hospitals securely.
Authors
Xiaofei Wu, Rongmei Liang, Jian Qing Shi
Abstract
Multimodal fusion is studied under hybrid federation, where data are partitioned across samples and feature blocks. The proposed two-dimensional federated smooth multiple-kernel large-margin learning method (2D-FedSMKL) integrates availability-corrected modality scores, group-kernel regularization, and smooth worst-site risk aggregation. For fixed maps and propensities, the objective is convex and smooth. Curvature bounds and exact synchronous block-gradient equivalence are established for binary and multiclass scores, giving a conditional accelerated rate. An oracle-propensity binary risk bound and a propensity-perturbation result are also derived. Simulations and two public activity-recognition benchmarks assess site heterogeneity and modality absence. Worst-site aggregation improves worst-site performance under heterogeneity, including reversed and size-balanced designs. On the benchmarks, validation selects the correction; the selected model is not significantly worse than the strongest reference at the worst site, and group-regularization benefits vary by dataset.