Abstract: Rating data of Internet applications contain abundant information about users' opinions and preferences. In order to discover user preference more accurately, this paper proposes a user preference modeling approach upon rating data based on Bayesian network, considering users' rating score and review data. First, the method for extracting different review attributes from the review data is presented. Then, the initial structure constraints and parameter constraints of the user preference model are selected from score and review. Finally, a user preference modeling method based on constraints is presented. Experimental results show that, compared with the user preference model constructed by score or review data alone, the user preference model which considers both score and review data can estimating user preference more accurately.