Matrix Factorization without User Data Retention

Recommender systems often rely on a centralized storage of user data and there is a growing concern about the misuse of the data. As recent studies have demonstrated, sensitive information could be inferred from user ratings stored by recommender systems. This work presents a novel semi-decentralized variant of the widely-used Matrix Factorization (MF) technique. The proposed approach eliminates the need for retaining user data, such that neither user ratings nor latent user vectors are stored by the recommender. Experimental evaluation indicates that the performance of the proposed approach is close to that of standard MF, and that the gap between the two diminishes as more user data becomes available. Our work paves the way to a new type of MF recommenders, which avoid user data retention, yet are capable of achieving accuracy similar to that of the state-of-the-art recommenders.