Federated Gaussian Process: Convergence, Automatic Personalization and Multi-fidelity Modeling.

Area of Science:

• Machine Learning
• Statistical Modeling
• Data Science

Background:

• Federated learning enables collaborative model training without sharing raw data.
• Gaussian Processes (GPs) are powerful non-parametric models for regression tasks.
• Personalized modeling requires capturing unique features from individual datasets.

Purpose of the Study:

• To propose FGPR, a Federated Gaussian Process regression framework.
• To enhance personalization in federated learning through shared priors.
• To provide theoretical guarantees for FGPR's convergence and demonstrate its practical utility.

Main Methods:

• Federated Gaussian Process (GP) regression (FGPR) framework.
• Model aggregation using an averaging strategy.
• Local computations via stochastic gradient descent.
• Joint learning of a shared prior across devices.

Main Results:

• FGPR achieves personalized predictions by combining a shared prior with local data.
• Theoretical analysis shows FGPR converges to a critical point of the log-marginal likelihood.
• Extensive case studies demonstrate FGPR's effectiveness in various regression tasks.
• FGPR proves to be a promising approach for privacy-preserving multi-fidelity data modeling.

Conclusions:

• FGPR offers a robust and privacy-preserving method for federated regression.
• The framework effectively balances global knowledge sharing with local data personalization.
• FGPR advances theoretical understanding of federated learning in correlated settings.