Considering rating as probability distribution of attitude in recommender system

Recommender systems play increasingly significant roles in solving the information explosion problem. Generally, the user ratings are treated as ground truth of their tastes, and used as index for later predict unknown ratings. However, researchers have found that users are inconsistent in giving their feedbacks, which can be considered as rating noise. Some researchers focus on improving recommendation quality by de-noising user feedbacks. In this paper, we try to improve recommendation quality in a different way. The rating inconsistency is considered as an inherent characteristic of user feedbacks. User rating is described by the probability distribution of user attitude instead of the exact attitude towards the current item. According to it, we propose a recommendation approach based on conventional user-based collaborative filtering using the Manhattan Distance to measure user similarities. Experiments on MovieLens dataset show the effectiveness of the proposed approach on both accuracy and diversity.