Aiming at the problems of sparse data, cold start, timeliness and privacy protection in current recommendation systems, a collaborative filtering recommendation algorithm based on density weight and improved similarity was proposed. The collaborative filtering recommendation algorithm, which combines differential privacy protection clustering and improved similarity, aims to improve the accuracy of the recommendation system and ensure the privacy security of user data. The user-project score matrix was constructed through data pre-processing, and the Weight Slope One algorithm was used to fill empty values in an intelligent way. The DWDPK-me-doids privacy clustering algorithm was used to cluster the matrix accurately, and the time factor and user interest preference factors were integrated to change the calculation of similarity, thus improving the relevance of recommendation. Finally, the target user's rating of the project was predicted. Comparative experiments were conducted on the MovieLens dataset against five privacy recommendation algorithms proposed by current scholars validate the efficacy of the proposed algorithm, showing reductions in evaluation metrics such as root mean squared error(RMSE) and mean absolute error(MAE). This indicates that the method partially addresses issues such as data sparsity, cold start, and timeliness, while enhancing recommendation accuracy on the basis of protecting user privacy.