Federated Deep Latent Factor Model for Privacy-Preserving Recommendation

Recommender systems (RSs) are extensively applied in various domains, such as e-commerce and online media services, to enhance user experience. Traditional RSs rely on centralized data collection and model training, which, while effective, raise serious privacy concerns as user interaction data may contain sensitive personal information. Federated Learning (FL) has emerged as a promising solution by allowing users to collaboratively train a global model without sharing their raw data. However, federated recommender systems face challenges, such as data sparsity, personalization limitations, and difficulties in capturing complex relationships between users and items. To address these challenges, we propose a novel federated deep latent factor framework (FedDeepLF) to improve recommendation accuracy while preserving user privacy. FedDeepLF introduces three key innovations: (1) utilizing deep neural networks to model complex user-item interactions; (2) incorporating user ratings into the model to better capture individual preferences; and (3) generating synthetic ratings to protect user interactions while improving recommendation quality. Extensive experiments on five datasets demonstrate that FedDeepLF significantly outperforms state-of-the-art federated recommender models in terms of prediction accuracy.