The key challenge of personalized federated learning (pFL) is to capture the statis
tical heterogeneity properties of data with inexpensive communications and gain
customized performance for devices. To address this, we introduced personalized
multi-tier federated learning with Moreau envelopes (pFedMT) to obtain optimized
and personalized local models when there are known cluster structures across
devices. Moreau envelopes are used as the devices’ and teams’ regularized loss
functions. We provide theoretical guarantees of pFedMT for both strongly convex
and non-convex settings. pFedMT achieves linear convergence rates for strongly
convex settings and sub-linear rates for non-convex and smooth settings.