The two-layer restricted Boltzmann machine (RBM) for rating prediction problems in recommendation systems has become one of the significant researches. A dual conditional restricted Boltzmann machine (dCRBM) model is proposed in this paper. In the dCRBM model, the training process uses rated/unrated information. Meanwhile, we also utilize the dual patterns of users and items, and build two CRBMs based on users and items respectively. Moreover, rating predictions of these two CRBMs via regression models are combined. The experimental results from MovieLens dataset show that the dCRBM model proposed in this paper helps to improve the prediction accuracy of the recommendation system.