Applying sparsity- and overfitting-aware eXtreme Gradient Boosting (XGBoost) for classification in federated learning allows many participants to train a series of trees collaboratively. Since various local multiclass distributions and global aggregation diversity, model performance plummets as convergence slowly and accuracy decreases. Worse still, neither the participants nor the server can detect this problem and make timely adjustments. In this paper, we provide a new local-global class imbalance inconsistency quantification and utilize softmax as the activation and focal loss, a dynamically scaled cross-entropy loss, in federated XGBoost to mitigate local class imbalance. Moreover, we propose a simple but effective hyperparameter determination strategy based on local data distribution to adjust the sample weights among non-communicating participants, synergetic focal loss, to solve the inconsistency of local and global class imbalance, a unique characteristic of federated learning. This strategy is perfectly integrated into the original classification algorithm. It requires no additional detectors or information transmission. Furthermore, a dynamical for loop is designed to capture an optimum hyperparameter combination. Finally, we conduct comprehensive tabular- and image-based experiments to show that synergetic focal loss used in federated XGBoost achieves faster convergency and significant accuracy improvement. Simulation results prove the effectiveness of the proposed principle of configuring sample weights.
In real-time systems, energy consumption is one of the most critical challenges. Dynamic voltage and frequency scaling (DVFS) algorithms have been widely applied to balance the trade-off between performance and power consumption. We first review some classic DVFS methods, e.g., LA-EDF and CCEDF, and then a hybrid DVFS algorithm (Soft-LA2) that we proposed recently. Soft-LA2 needs domain knowledge to set the trade-off parameter manually to achieve energy saving compared with classic methods. Motivated by recent deep learning technologies applied in DVFS, in this paper, we propose a new framework for Soft-LA2. Our method automatically optimizes parameters for Soft-LA2, which leads to more power saving compared with random parameter setting in Soft-LA2. Simulation results show that under certain neural network architecture setting, our method can find the best parameters automatically to achieve power saving performance. Furthermore, with test data, our proposed method saves 4.9% and 9.6% power consumption compared with relevant learning method DQL-EES and random setting Soft-LA2, respectively.
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.