Recently, collaborative tagging has been gaining increasing popularity in a large variety of websites. The tags generated from collaborative tagging provide highly abstracted information about users' personal taste on information item, and therefore could be used to profile both users and items. However, flattening the three dimensional user-tag-item matrix into two-way matrices will lead to loss of two dimensional relationships between users, items and tags completely. In this paper, we use a predictive bilinear model to capture the informative interaction patterns among user-tag and item-tag matrices, and employ the Nonnegative Matrix Factorization (NMF) algorithm to extract lower dimensional representative features from tags. Experiments on two real-world datasets show that our approach substantially outperforms the traditional CF methods as well as tag-based recommendation methods reported in the literature.