Federating recommendations using differentially private prototypes

Machine learning methods exploit similarities in users’ activity patterns to provide recommendations applications across a wide range of fields including entertainment, dating, and commerce. However, domains that demand protection personally sensitive data, such as medicine or banking, how can we learn recommendation models without accessing the data inadvertently leaking private information? Many situations medical field prohibit centralizing from different hospitals thus require information kept separate databases. We propose new federated approach global local for collecting raw user statistics, about personal preferences. Our method produces set locally learned prototypes allow us infer behavioral while providing differential privacy guarantees users any database system. By requiring only two rounds communication, both reduce communication costs avoid excessive loss associated with typical iterative procedures. test our framework on synthetic real version Movielens ratings. show adaptation model allows proposed outperform centralized matrix-factorization-based recommender system models, terms accuracy matrix reconstruction relevance recommendations, maintaining provable guarantees. also is more robust has smaller variance than individual by independent entities.