USST: a two-phase privacy-preserving framework for personalized recommendation with semi-distributed training

Personalized recommendations are becoming indispensable for assisting online users in discovering items of interest. However, existing recommendation algorithms rely heavily on the collection of personal information, which poses significant privacy concerns to users. In this paper, we propose a two-phase privacy-preserving framework called user sampling and semi-distributed training (USST) for personalized recommendations, which can protect user privacy while ensuring high recommendation accuracy. In the USST framework, rather than directly training the model with all user records, a shared model is first trained with a small set of records contributed by sampled users (e.g., paid users and volunteers). This shared model is then distributed to each user, who further trains a personalized model using personal information. Thus, the USST guarantees that all unsampled users never disclose their private information. To validate the effectiveness and practicality of USST, we designed two USST-based privacy-preserving recommendation algorithms, USST-SVD and USST-NCF based on SVD and NCF algorithms, respectively. We conducted evaluations using MovieLens and Netflix Prize datasets, and the results show that, using only 20% of sampled users’ records, the recommendation accuracy of USST-based algorithms is very close to that of all users’ records. Thus, USST can significantly improve the level of privacy protection in recommender systems.