Collaborative filtering-based models can use the interaction between users and products or the correlation between users and users, and between products and products. However, methods based on collaborative filtering can only grasp one type of relationship and still cannot fully fit. Various factors influencing user preferences make a lot of redundant information still not filtered out. We proposals a collaborative filtering model based on deep learning, which combines the item-item relationship learning in advance with a neural collaborative filtering network to effectually make recommendations. In the initial stage, learn low-dimensional vectors of compartments, and embed information that reflections the co-occurrence relationship between compartments. The prediction stage combines the trained embedding vector with the embedding vector of the module as a correction to the output result of the neural network. The benchmark data set MovieLens 1M is the experienced data set of this article, and the effectiveness of this method is verified on the data set. The experienced results are compared with some advanced methods on the data set. The results show that the model proposed in this paper is better than some methods based on collaborative filtering.