Federated learning (FL) trains global models across distributed devices without sharing local data; however, it is confronted by significant challenges arising from non-Independent and Identically Distributed (Non-IID) data and the absence of effective incentive mechanisms. Both factors impede client participation and degrade the overall model performance. To address these challenges, we propose a novel personalized FL framework integrating hierarchical Bayesian modeling and variational expectation-maximization. By modeling the global parameter as a latent variable, our approach captures shared client trends while preserving individual heterogeneity. To quantify contributions, we introduce a head-based architecture that employs confidence-based personalized metrics and gradient cosine similarity to achieve global alignment, ensuring the fairness and efficiency of reward distribution. Experimental results on the MNIST and CIFAR-10 datasets, under varying degrees of data heterogeneity, consistently demonstrate that our method outperforms four widely adopted baseline algorithms in terms of model accuracy, contribution evaluation, and fairness. This framework lays a strong practical foundation for encouraging active participation in FL systems, especially in the face of highly heterogeneous data distributions.