Personalized federated learning improves privacy and prediction accuracy
Personalized Federated Hierarchical Gaussian Processes for Privacy-Preserving Modeling of Heterogeneous Distributed Systems
Machine Learning
Summary
Data is often spread across many different users or devices, each with their own unique patterns. The authors propose a way to share knowledge between these users to improve predictions without sharing their private data directly. Their method breaks down data patterns into shared global trends, individual differences, and unique local parts, making the model both personalized and privacy-preserving. Their approach can accurately classify faults in manufacturing and identify pollution zones from multiple air quality stations without collecting all the data in one place.
What this means in practice
- •For manufacturing quality teams: Detect faults in industrial equipment using limited labeled cycles while keeping sensor data private within plants.
- •For environmental monitoring teams: Map geographic pollution zones by integrating distributed air quality sensors without centralizing time series data.
Authors
Xianjian Xie, Hao Yan
Abstract
We present Personalized Federated Hierarchical Gaussian Processes (pFedHGP) for probabilistic regression and classification when data are distributed across heterogeneous clients. Each client's latent function decomposes into (i) a shared global component, (ii) a client-specific deviation that shares the global kernel structure, and (iii) a flexible local residual. Sparse inducing-variable approximations and federated variational inference keep raw data local while the server synchronizes only low-dimensional statistics for the shared component. Full predictive distributions support uncertainty-aware decisions. In application studies, pFedHGP attains perfect fault classification in press tonnage monitoring using 13.77% of labeled cycles and recovers geographic zones in federated air-quality modeling without centralizing station-level time series. An Instantaneous Linear Mixing Model viewpoint links the hierarchy to multi-output Gaussian processes for correlated sensors.