P-Fed Rec: A Certifiable Unlearning Framework for Personalized Federated Recommendation

With the enactment of global data privacy regulations (e.g., GDPR), traditional centralized personalized recommendation systems face severe compliance challenges. How to maintain the personalized performance of recommendation systems while satisfying strict privacy regulations has become a core issue to be urgently addressed. Federated learning provides a new paradigm for privacy-preserving recommendations by keeping data locally trained. However, existing federated recommendation systems generally ignore users' "Right to be Forgotten" and struggle to achieve effective personalization and verifiable unlearning on highly non-independent and identically distributed (non-IID) user data. This paper proposes a novel federated recommendation framework, P-FedRec, to resolve the core contradiction between privacy compliance and model utility. The framework includes three major innovations: 1) an influence sampling strategy based on gradient hashing to record key trajectories of model evolution with extremely low storage overhead; 2) a certifiable unlearning algorithm with bias correction, which can efficiently and accurately remove specific users' contributions from the global model while minimizing interference with the overall system performance; 3) a set of verifiability proofs based on Lipschitz continuity, providing theoretical guarantees for unlearning operations. Experiments on the MovieLens-1M and Amazon-Book datasets show that compared with existing methods, P-FedRec achieves an average improvement of 2.3% in recommendation accuracy (NDCG@10), an unlearning efficiency improvement of more than three orders of magnitude, while maintaining a performance degradation within 1.8%. This framework provides a feasible technical path for building next-generation recommendation systems that comply with global privacy regulations and offers a new full-lifecycle management perspective for federated learning system design.