In view of performance flaws in the case of sparse data and the lack of similarity measurement methods in traditional collaborative filtering recommendation algorithm, a collaborative filtering recommendation algorithm based on multi-level hybrid similarity was proposed to improve the recommendation accuracy. The algorithm is divided into three different levels. Firstly, the concept of fuzzy set was used to fuzzify the user rating and then to calculate the user's fuzzy preferences, and the adjusted cosine-based similarity of the user rating and the Jarccad similarity of the user rating were combined as the user rating similarity. Secondly, the use rating was classified to predict the degree of interest of the user to the item category so that the user's interest similarity was calculated. Thirdly, the user's characteristic similarity was predicted by the characteristic attributes between users. Then, the user's interest similarity and user's characteristic similarity were dynamically integrated by the number of user ratings. Finally, the similarities of three levels were fused as the result of user similarity. The experimental results show that the improved hybrid algorithm has a decrease of 5% in Mean Absolute Error (MAE) compared to the adjusted cosine-based similarity algorithm when the number of neighbors is small. Compared with the improved MKJCF (Modified K-pow Jaccard similarity Cooperative Filtering) algorithm, the improved hybrid algorithm has a slight advantage, and the MAE fell by an average of about 1% with the increase of neighbor number. The proposed algorithm uses a multi-level recommendation strategy to improve the user's recommendation accuracy, effectively alleviates the sparseness of data and the impact of single measurement method.