The data sparsity and cold start problem of collaborative filtering recommendation algorithm affects and constrains the quality of recommendation. Trust calculation based on user-project bipartite graph can effectively utilize the hidden connection between users. The algorithm model proposed in this study includes the calculation of improved user-project adaptive trust, incorporates an improved enhanced trust mechanism based on user preferences, linear weighted JMSD correlation coefficient. Experiments under the two sets of data set show that the improved algorithm model has a lower Mean Absolute Error (MAE) and higher recall rate (Recall) than the three benchmark algorithms, both models have their own focus, which improves the quality of recommendations.