Like other recommendation paradigms, the algorithm based on collaborative filtering also suffers from the curse of sparse data and cold-start problem. Nevertheless, a recommendation algorithm based on collaborative filtering and hybrid similarity model was proposed. First, the similarity of users among different items was calculated by the algorithm, and then the relationship among users, items, features and tags was described according to property weights and tag weights. Next, the score preference between different users was adjusted by setting the user preference factor and the asymmetry factor. After that, a hybrid similarity model was constructed based on the user similarity, item weights and rating preferences, and the user-time weight information was also added to solve the project cold-start problem. Experiments on publicly accessible MovieLens data sets demonstrate that the algorithm achieves more prominent results than other related approaches in various evaluation metrics.