Recommender systems use ratings from users on items such as movies and music for the purpose of predicting the user preferences on items that have not been rated. Predictions are normally done by using the ratings of other users of the system, by learning the user preference as a function of the features of the items or by a combination of both these methods. In this paper, we pose the problem as one of collaboratively learning of preference functions by multiple users of the recommender system. We study several mixture models for this task. We show, via theoretical analyses and experiments on a movie rating database, how the models can be designed to overcome common problems in recommender systems including the new user problem, the recurring startup problem, the sparse rating problem and the scaling problem. 1.