Skip to Main Content
With the explosive growth of online social communities and massive user-generated content, privacy-preserving recommender systems, which identify information of interest to individual users without disclosing personal interests to other parties, have become increasingly important. Collaborative filtering (CF), a widely used recommendation technique, recommends content that similar users have liked. As a result, CF-based recommender systems may expose sensitive personal interest information. This is demonstrated by a privacy attack model we present that targets online social communities. To solve this problem, we propose an interest group based privacy-preserving recommender system called Pistis. By identifying inherent item-user interest groups and separating users' private interests from their public interests, Pistis can make recommendations based on aggregated judgments of group members and local personalization, thus avoiding the disclosure of personal interest information. Pistis has been deployed and evaluated in an online social community with over 63,000 users, 20,000 daily posts, and 180,000 daily reads. Compared with two representative CF-based methods, our evaluation results demonstrate that Pistis achieves better performance in privacy preservation, recommendation quality, and efficiency.