The sparsity of user-item ratings is a common problem in collaborative filtering recommender systems. In traditional collaborative filtering recommender systems, similarity of users is often calculated with cosine and Pearson methods based on common ratings. When user-item ratings are sparse, the ratio of common rated items is less, and the accuracy of recommendations will be influenced because users with similar preferences can't be found accurately. To change calculation method of user similarity based on the same rated items, this paper applies EMD (earth mover's distance) to implement cross-item similarity calculation of mobile user and proposes a collaborative filtering recommendation method combining item features and trust relationship of mobile users. The experimental results show that, comparing with cosine and Pearson, user similarity calculation method combining item features can relieve influence of the sparsity of user-item ratings on collaborative filtering recommender systems. And the proposed recommender method can improve accuracy of mobile recommendations.