FedREPA: Relation-Enhanced Prototype Alignment for Heterogeneous Federated Human Activity Recognition

Federated human activity recognition (HAR) with wearable sensors is challenged by label-skewed client data, missing local activities, and scarce mini-batch negatives. We propose FedREPA, a relation-guided framework that couples missing-class-aware relation consistency with prototype contrastive learning. Clients derive an inter-class relation prior from shared global class prototypes. A masked relation-consistency loss matches supported local relation rows to this prior. The same prior reweights negative classes in the prototype contrastive loss, giving greater emphasis to nearby prototypes when trust is held fixed. Global prototypes provide cross-client positive and negative anchors, including negatives for activities absent locally. On UCI HAR, PAMAP2, CIFAR-10, and CIFAR-100 under simulated Dirichlet label skew, FedREPA records the highest observed mean global test accuracy in the comparison with the ten evaluated baselines at each tested Dirichlet level, using the last five evaluation points per run. FedTGP, FedDPA, and FedAli-PHAR were not evaluated, and no performance advantage over them is claimed. Its paired accuracy margin over the restricted FedGPS-Core baseline averages about one percentage point. Worst-client accuracy is also highest among the methods shown on UCI HAR and CIFAR-10, while selected-class recall on CIFAR-10 shows mixed results. These findings are descriptive comparisons under the evaluated Dirichlet label-skew protocol. Performance under subject-wise federated partitions, device-wise heterogeneity, and sensor-placement variation remains unvalidated, as does clinical applicability.