Abstract: A novel bi-level similarity collaborative filtering (BLSCF) algorithm is proposed in this paper to deal with the problem of low recommendation precision and accuracy of collaborative filtering algorithms resulted from the data sparsity of user evaluation matrix. Different from the existing algorithms which improve the recommendation precision and accuracy via modifying the similarity or mixed similarity, this paper distinguishes the nearest neighborhood similarity from the nearest score similarity and utilizes the bi-level similarity to search their neighbors. The first level is to obtain the nearest neighbors of the user's behavior preference by integrating the log-likelihood ratio of the user's common and difference scoring behaviors with the user's preference similarity on item attributes. The second one will search the nearest rating neighbors and measure the similarity of the user's score via the improved Pearson similarity so as to rate and predict the user's unknown items. It is shown via experimental results on the Movielens dataset that the proposed BLSCF algorithm in this work can rapidly eliminate the interference to find the user's neighbor and greatly improve the accuracy and precision of the recommendation system.