Adaptive Personalized Federated Learning via Multi-task Averaging of Kernel Mean Embeddings

Researchers propose a new Personalized Federated Learning approach that automatically learns optimal collaboration weights between agents without prior knowledge of data heterogeneity. The method uses kernel mean embedding estimation to capture statistical relationships between agents and includes a practical implementation for communication-constrained federated settings.

• →The approach formulates collaborative weight estimation as a kernel mean embedding problem with multiple data sources.
• →The method automatically transitions between global and local learning regimes without requiring prior knowledge of data heterogeneity.
• →Finite-sample guarantees on local excess risks are provided for a broad class of distributions.
• →A practical implementation using random Fourier features addresses communication constraints in federated settings.
• →The approach allows trading communication cost for statistical efficiency in collaborative learning.