A Differentially Private SVD Recommendation Algorithm with User Tags

With the continuous expansion of Internet commerce and the rapid increase in both amount and types of products, recommendation algorithms play an important role in selecting items of interest to users. Although the classic collaborative filtering recommendation algorithms are simple in principle and has low time complexity, it fails to make full use of the user tags generated by users, which may contain implicit user preferences, resulting in low prediction accuracy and cold start problems. Besides, the classic collaborative filtering algorithms don’t take privacy disclosure issues into consideration, and may expose user preferences during the recommendation process. In order to solve the above problems, a recommendation algorithm named DP-SRAUT (Differentially Private SVD Recommendation Algorithm with User Tags) is presented in the paper. The DP-SRAUT algorithm firstly pre-processes the user's evaluation tag set of items, then generates the users’ semantic preference matrix for items, and the biased SVD approach is used for prediction and item recommendation with the integration of the users’ semantic preference matrix. To preserving user privacy, Laplace noise is added during the training process to meet ε -difference privacy. Experiment results on the Movielens dataset show that the proposed algorithm achieves better accuracy in condition of keeping user privacy.