Collaborative ltering-based recommender systems exploit user preferences about items to provide them with recommendations. These preferences are generally ratings. However, choosing a rating is no easy task for any user; the rating scale is usually reduced and the rating values given by the users may be in uenced by many factors. The rat- ings are thus not completely trustworthy. This paper is a rst attempt at studying the expression of preferences in collaborative ltering under the form of preference relations instead of ratings. When using preference relations, users are asked to compare pairs of resources. We propose new measures to compute recommendations using preference relations. First experiments have been conducted on a state of the art corpus of the rec- ommender systems domain and show that this new approach compares with, and in some cases improves the classical one.