Rating prediction is an important problem for rating based recommender systems. In Rating Prediction, the task is to predict the rating that a user would give to an item that he/she has not rated in the past. Most of the existing algorithms for the task concentrate on the absolute ratings given to different items by different users in the past. However, there are few recent research work that point out some drawbacks of absolute rating based systems and algorithms, and suggest the use of preference relations between pairs of items to capture the users' interests about the items. In this paper, we propose a rating prediction algorithm that considers the relative ratings given by the users for different pairs of items. The algorithm models the users and items using a matrix factorization framework. The learned model of users and items are first used to predict the personalized utility of an item for a user. This utility is then converted to a valid rating value in a predefined rating scale by employing a personalized scaling. Experimental evaluation on a benchmark dataset reveals that better prediction accuracies may be achieved by modeling the users and items using relative rating information.