Energy-Efficient Personalized Federated Learning via Probabilistic Global–Local Model Updates

Personalized federated learning (PFL) is essential for mobile sensing with user-dependent input–output semantics, but existing methods rarely tie personalization demand to communication-aware scheduling that trades off costly uplink transmission against on-device computation under battery and time-varying wireless constraints. This paper bridges the systems–learning gap by making personalization demand explicit and using it to jointly schedule uplink communication and on-device computation. We introduce an output-level global–local mixing architecture in which a learnable mixing coefficient explicitly captures each client’s reliance on shared versus local predictions, rendering the value of global coupling measurable and hence schedulable. We propose ROFL (Ratio-Optimized Personalized Federated Learning), a branch-wise three-action probabilistic scheduling scheme: in each round, a client either updates the global branch with aggregation, updates the local branch only, or skips to conserve energy. Although each round updates only one branch under nonuniform action sampling, we establish a nonconvex convergence bound for this partial-update process, yielding an explicit energy–accuracy characterization; it further induces a closed-form local-to-global scheduling ratio governed by personalization level and the communication–computation cost regime. Experiments on CIFAR-10 with controlled label/semantic heterogeneity and on a real-world smartphone stress monitoring dataset show that ROFL consistently improves energy-to-target efficiency under moderate-to-severe heterogeneity, reducing the energy to reach the same target accuracy by up to ∼75% versus representative FL/PFL baselines, while remaining competitive in near-i.i.d. regimes.