Prediction uncertainty in collaborative filtering: Enhancing personalized online product ranking

• We propose a novel approach (RPU) for personalized online product ranking.
• RPU improves the accuracy of ranking by considering prediction uncertainty.
• Posterior distribution and confidence level are used as key factors for uncertainty.
• Experiments using real-world data show that RPU achieves advantageous performance.
• The results are robust in terms of sparse data.

Personalized product ranking provides support to the decision making of online consumers and helps improve their satisfaction, since consumers always face a large volume of choices when they are shopping online. Recommender systems with collaborative filtering techniques are commonly used for this purpose, wherein products are ranked according to their predicted ratings. However, this kind of ranking approaches (namely, Ranking by Collaborative Filtering, RCF for short) have generally ignored the impacts of prediction uncertainty. This paper proposes a novel ranking approach called RPU (Ranking with Prediction Uncertainty), which utilizes posterior rating distribution and confidence level of prediction as two key factors for prediction uncertainty. Serving as a critical component of the generalized ranking framework, RPU aims to improve the accuracy of personalized product ranking through incorporating the uncertainty information. Experiments using real-world data of movie ratings show that RPU achieves higher ranking performance compared to traditional RCF and the results are robust in terms of sparse data.