Hierarchical privacy-preserving federated recommendation via adaptive multi-level personalization and dynamic aggregation

Federated recommendation systems aim to balance personalization and privacy protection, but this is extremely challenging in heterogeneous client environments. Current methods, while offering some personalization, fail to capture the inherent hierarchical structure in user preferences—a combination of individual taste, community influence, and global trends—nor can they effectively adapt to data differences between clients. To this end, we propose HierFedRec, an innovative hierarchical federated recommendation framework. The framework introduces three core mechanisms: (1) a three-tier embedding architecture that characterizes items from individual, cluster, and global dimensions, and adaptively fuses these perspectives for deep personalization; (2) privacy-aware adaptive learning, which combines differential privacy with dynamic learning rates to handle client heterogeneity; and (3) quality-driven aggregation, which determines the weight of model updates based on data quality and inter-client similarity, rather than simple averaging. Our framework efficiently manages the privacy budget by injecting noise only once at the server side and mitigates performance degradation from client drift through adaptive learning rates. Extensive experiments on six real-world datasets demonstrate that HierFedRec significantly outperforms existing state-of-the-art methods on both HR@10 and NDCG@10 metrics, while providing rigorous (ϵ,δ)-differential privacy guarantees. Analysis reveals that our hierarchical structure, particularly the cluster-level embeddings, serves as a crucial bridge connecting individual preferences with global patterns, which greatly enhances recommendation performance for cold-start users. HierFedRec offers a clear path for building practical, efficient, and secure personalized recommendation systems in cross-device scenarios.