In recent years, deep neural network is introduced in recommender systems to solve the collaborative filtering problem, which has achieved immense success on computer vision, speech recognition and natural language processing. On one hand, deep neural network can be used to model the auxiliary information in recommender systems. On the other hand, it is also capable of modeling nonlinear relationships between users and items. One advantage of deep neural network is that the performance of the algorithm can be easily enhanced by augmenting the depth of the neural network. However, two potential problems may emerge when the deep neural work is exploited to model relationships between users and items. The fundamental problem is that the complexity of the algorithm grows significantly with the increment in the depth of the neural network. The second one is that a deeper neural network may undermine the accuracy of the algorithm. In order to alleviate these problems, we propose a hybrid neural network that combines heterogeneous neural networks with different structures. The experimental results on real datasets reveal that our method is superior to the state-of-the-art methods in terms of the item ranking.