As a key solution to the problem of information overload, the recommender system can filter a large deal of information according to user’s preference and provide personalized recommendations for the user. However, traditional collaborative filtering models with excellent performance haven’t made full use of the contextual information in the process of recommendation, which to some extent confronts the system with the performance bottleneck. In order to improve the system performance further, this paper starts with the contextual information on ratings, and proposes a collaborative filtering model fusing singularity and diffusion process (CFSDP) by taking advantage of ratings’ singularities obtained from the classified statistics of ratings and referring to the similarity model of multi-channel diffusion which regards recommender system as a user-item bipartite network. To demonstrate the superiority of the proposed model, the study provides comparative experimental results based on the MovieLens, NetFlix and Jester data sets. Finally, the results show that the model not only has better extensibility, but also can observably improve the prediction and recommendation quality of system with a reasonable time cost.