Privacy-Preserving Collaborative Filtering Based on Time-Drifting Characteristic

Recommendation has become increasingly important because of the information overload. Collaborative filtering (CF) technique, as the most popular recommendation method, utilizes the historical preferences of users to predict their future interests on other items. However, CF technique requires collecting users' rating information, which may lead to the disclosure of privacy. We propose a new randomized perturbation approach Time-drifting privacy-preserving collaborative filtering (TPPCF) to well balance privacy of users and accuracy of recommendation. Since users' recent ratings can better represent their interests and preferences, we incorporate a varying weight into the approach. Specifically, we assign higher weights to more recent ratings both when computing user similarity and perturbing users' ratings. To further improve the efficiency, we cluster the users into several groups to reduce computation cost. We demonstrate the effectiveness and efficiency of our method through experiments on MovieLens dataset, which shows TPPCF can achieve higher privacy while generating more accurate recommendation.