The relation between users and items in the recommendation system was considered as a complex network structure, and the binary relation between users and items was used to construct a graph model,based on which a diffusion dynamics method was introduced to study the recommendation algorithm.Through studying the ideas spreading in the network structure and collaborative filtering algorithms,two optimized algorithms were presented based on user similarity and item similarity,respectively.The proposed approach provides a correction method for zero value problems of similarity calculation,ignored in most existing publications.The similarity of users (or items) was defined as the number of corresponding views which a user (or items) owns and differences between those viewpoints spreading in the complex network.Using MovieLens data set,the experiments show that the presented algorithm has better performance than the collaborative filtering algorithm based on Pearson correlation coefficient and some other existing methods.