Learning individual preferences of users for items from user-item rating matrix is critical for rating recommendation. Many recommendation methods, such as the Latent Factor model, can not make full use of the interaction information from rating matrix to learning individual preferences, and achieve unsatisfying results. Inspired by wide and deep learning model of deep learning in APP recommendation, deep hybrid model is proposed and named DeepHM for rating recommendation. Compared with the wide and deep model, deep wide model and DNN model are used to reconstruct wide model and deep model, which can get DeepHM and make DeepHM become shared input to its deep wide and deep parts. Therefore, DeepHM uses interaction information of user and item from rating matrix more efficiently to obtain individual preferences information. Furthermore, DeepHM treats the rating recommendation as a multi-classification problem aiming to improve the accuracy of recommendation. Through comprehensive experiments on public Movielens datasets, it demonstrates that the efficiency of DeepHM based on rating recommendation is better than that of the existing models.