Personalized Privacy Preserving Collaborative Filtering

Recommendation systems are widely applied these years as a result of significant growth in the amount of online information. To provide accurate recommendation, a great deal of personal information are collected, which gives rise to privacy concerns for many individuals. Differential privacy is a well accepted technique for providing a strong privacy guarantee. However, traditional differential privacy can only preserve privacy at a uniform level for all users. When, in reality, different people have different privacy requirements. A uniform privacy standard cannot preserve enough privacy for users with a strong privacy requirement and will likely provide unnecessary protection for users who do not care about the disclosure of their personal information. In this paper, we propose a personalized privacy preserving collaborative filtering method that considers an individual’s privacy preferences to overcome this problem. A Johnson Lindenstrauss transform is introduced to pre-process the original dataset to improve the quality of the selected neighbours - an important factor for final prediction. Our method was tested on two real-world datasets. Extensive experiments prove that our method maintains more utility while guaranteeing privacy.