Social network sites have become an important medium for people to receive information anytime anywhere. Users of social network sites share information by posting updates. The updates shared by friends form social update streams that provide people with up-to-date information. To receive novel information, users of social network sites are encouraged to establish social relations. However, having too many friends can lead to an information overload problem causing users to be overwhelmed by the huge number of updates shared continuously by numerous friends. The information overload problem can result in bad user experiences. It may also affect user intentions to join social network sites and thereby possibly reduce the sites’ advertising earnings which are based on the number of users. To resolve this problem, there is an urgent need of effective friend recommendation methods. A user is considered as a valuable friend if people like the updates the user posts. In this paper, we propose a model-based recommendation method which suggests valuable friends to users. Techniques of matrix factorization and learning to rank are designed to model the latent preferences of users and updates. At the same time, social influence is incorporated into the proposed method to enhance the learned preferences. Valuable friends are recommended if the preferences of the updates that they share are highly associated with the preferences of a target user. Our experiment findings that are based on a huge real-world dataset demonstrate the effectiveness of the social influence and learning to rank on a friend recommendation task. The results show that the proposed method is effective and it outperforms many well-known friend recommendation methods in terms of the coverage rate and ranking performance.