The application of bipartite graph theory in collaborative filtering recommendation based on substance diffusion theory of complex networks has attracted more and more attention from scholars. Existing algorithms mainly consider the positive rating when calculating neighbor users, ignoring the negative rating of users. In order to improve the accuracy of recommendation algorithm, a collaborative filtering recommendation algorithm based on improved bipartite graph and user reliability is proposed. The algorithm quantifies both positive ratings and negative ratings into the weight of the path, which controls the user's energy distribution, and takes users' reliability into account when predicting the rating, therefore, the accuracy of recommendation result is significantly improved. A series of comparative experiments are carried out on MovieLens and Eachmove datasets. The experimental results show that the improved algorithm has lower mean absolute error.