Local Differentially Private Matrix Factorization For Recommendations

In recent years recommendation systems have become popular in the e-commerce industry as they can be used to provide a personalized experience to users. However, performing analytics on users’ private information has also raised privacy concerns. Therefore, various privacy protection mechanisms have been proposed for recommendation systems. Yet most of these methods provide privacy protection against user-side adversaries and disregards the privacy violations caused by the service providers. In this paper, we propose a local differential privacy mechanism for matrix factorization based recommendation systems. In the proposed method, users perturb their ratings locally on their devices using Laplace and randomized response mechanisms and send the perturbed ratings to the service provider. We evaluate the proposed mechanism using Movielens dataset and demonstrate that it can achieve a satisfactory tradeoff between data utility and user privacy.