An Efficient Federated Learning Approach for Dynamic User Personalization in Smart Digital Cities Under Data Heterogeneity

Federated Learning (FL) has emerged as a promising paradigm for privacy-preserving training of machine learning models in smart digital cities. However, it faces significant challenges, including data heterogeneity and dynamic user personalization requirements in real-world scenarios. Although personalized Federated Learning aims to address these issues, existing methods often exhibit notable limitations in handling non-independent and identically distributed (non-IID) data and adapting to diverse user needs. This paper proposes an efficient personalized federated learning framework that incorporates time-domain prioritization and data heterogeneity-aware mechanisms. Our approach leverages rational client grouping, dynamic sampling probability adjustment, and intra-group metalearning training to significantly enhance the performance and personalization capabilities of FL under non-IID conditions. Specifically, the client grouping strategy is designed based on cosine similarity and the K-means clustering algorithm to optimize grouping efficiency. In addition, the task prioritization problem is reformulated as a sampling probability optimization problem, which is addressed by a priority sampling algorithm. To further improve personalized performance, we introduce intragroup meta-learning training and inter-group aggregation, ensuring robust and adaptive model updates. Experimental results demonstrate that the proposed approach outperforms existing methods in terms of model accuracy and clustering performance, achieving significant improvements in both personalization and overall system efficiency.