There are data sparsity and scalability problems in collaborative filtering algorithms. Due to the huge amount of user and project data, the data is very sparse, and the data under different data sets are different, which makes the similarity calculation and user clustering effect in the existing algorithms not accurate enough and has a significant impact on the accuracy of the recommendation algorithms. In order to improve the accuracy of similarity calculation and nearest neighbor search, this paper proposes a collaborative filtering algorithm based on similarity optimization and manifold learning. By using the weighted factor to calculate the optimization similarity and combining the manifold learning, spectral clustering is carried out on sparse users after their scores are evaluated and the dimensionality is reduced. The obtained global optimal solution can improve the accuracy of the nearest neighbor of the target users in clustering, thus improving the recommendation accuracy of collaborative filtering. Experimental results on Epinions databset and Movielens databset show that the proposed algorithm can effectively reduce the mean absolute error and root mean square error of collaborative filtering algorithms, improve the recall rate and achieve higher recommendation accuracy.