A Local-Clustering-Based Personalized Differential Privacy Framework for User-Based Collaborative Filtering

The Collaborative Filtering (CF) algorithm plays an essential role in recommender systems. However, the CF algorithm relies on the user’s direct information to provide good recommendations, which may cause major privacy issues. To address these problems, Differential Privacy (DP) has been introduced into CF recommendation algorithms. In this paper, we propose a novel framework called Local-clustering-based Personalized Differential Privacy (LPDP) as an extension of DP. In LPDP, we take the privacy requirements specified at the item-level into consideration instead of employing the same level of privacy guarantees for all users. Moreover, we introduce a local-similarity-based item clustering process into the LPDP scheme, which leads to the result that any items within the same local cluster are hidden. We conduct a theoretical analysis of the privacy guarantees provided within the proposed LPDP scheme. We experimentally evaluate the LPDP scheme on real datasets and demonstrate the superior performance in recommendation quality.