Abstract
Collaborative fltering is now successfully applied to recommender systems. The availability of extensive personal data is necessary for generating high quality recommendations. However, traditional collaborative fltering methods suffer from sparse and sometimes cold-start problems, particularly for newly deployed recommenders. Currently, several recommender systems exist in good working order, and data collected from these existing systems should be valuable for newly deployed recommenders. This paper introduces a privacy preserving shared collaborative fltering problem in order to leverage the data from other parties (contributors) to improve its own (benefciaries) collaborative fltering performance, with the privacy protected under a differential privacy framework. It proposes a two-step methodology to solve this problem. First, item-based neighborhood information is selected as the shared data from the contributor with guaranteed differential privacy, and a practical enforcement mechanism for differential privacy is proposed. Second, two novel algorithms are developed to enable the beneficiary to leverage the shared data to support improved collaborative fltering. The extensive experimental results show that the proposed methodology can increase the recommendation accuracy of the benefciary significantly while preserving data privacy for the contributors.