Differentially private recommender systems

摘要: We consider the problem of producing recommendations from collective user behavior while simultaneously providing guarantees privacy for these users. Specifically, we Netflix Prize data set, and its leading algorithms, adapted to framework differential privacy.Unlike prior work concerned with cryptographically securing computation recommendations, constrains a in way that precludes any inference about underlying records output. Such algorithms necessarily introduce uncertainty--i.e., noise--to computations, trading accuracy privacy.We find several approaches competition can be provide privacy, without significantly degrading their accuracy. To adapt explicitly factor them into two parts, an aggregation/learning phase performed guarantees, individual recommendation uses learned correlations individual's personalized recommendations. The adaptations are non-trivial, involve both careful analysis per-record sensitivity calibrate noise, as well new post-processing steps mitigate impact this noise.We measure empirical trade-off between adaptations, non-trivial formal still outperforming Cinematch baseline provides.