Exploring the impact of LSH Parameters in Privacy-Preserving Personalization

The "Privacy versus Personalization" dilemma refers to the situation where in order to benefit from collaborative personalized service the users have to disclose their sensitive personal data. Solving this dilemma is challenging because generating collaborative filtering recommendations requires access to the set of all user profiles in order to identify similar ones, and to compute their top-rated items. The P3 paradigm (Privacy-Preserving Personalization) builds on the idea of using the Locality Sensitive Hashing (LSH) to find groups of similar users while keeping their profiles local. In this work, we analyze the behavior of the adapted LSH algorithm from the perspective of final recommendations quality and cluster size distribution. We investigate the impact of different LSH parameter configurations on the basis of the MovieLens dataset, and empirically show a small non-prohibitive cost of privacy protection on the recommendations quality.