Dual-view personalized federated recommendation

With the widespread adoption of federated learning (FL) in recommendation systems, balancing user privacy protection and personalized recommendation effectiveness has become a critical challenge. Previous federated recommendation models often mechanically apply traditional recommendation models to the federated setting, leading to item embeddings that are insensitive to specific client preferences, thereby compromising model performance. To address this issue, this paper proposes an intra-client personalization view designed to encode the diverse interests of clients toward the same item, ensuring that item representations are customized for different users to better reflect individual preferences. Furthermore, we introduce cross-client contrastive views to enhance client-side personalization, revealing differences in item representations across clients. Considering the varying impacts of these contrastive views on the final recommendation results, we incorporate an adaptive fusion mechanism that dynamically adjusts their weight contributions within the model to optimize recommendation performance. Experimental results demonstrate that DPFedRec significantly outperforms state-of-the-art personalized federated recommendation algorithms on three real-world datasets, validating its effectiveness in improving model performance, maintaining personalization, and preserving user privacy.