Aiming at the shortages in similarity calculation and rating prediction in traditional recommendation system, in order to further improve the accuracy and stability of the algorithm, a new recommendation algorithm was proposed. Firstly, according to the number of important labels for an item, the M ² similarity between the item and other items was calculated, which was used to constitute the nearest item set of the item. Then, according to the Slope One weighting theory, a new rating prediction method was designed to predict users' ratings based on the nearest item set. To validate the accuracy and stability of the proposed algorithm, comparison experiments with the traditional recommendation algorithms including K-Nearest Neighbor (KNN) algorithm based on Manhattan distance were conducted on MovieLens dataset. The experimental results showed that compared with the KNN algorithm, the mean absolute error and the root mean square error of the new algorithm were decreased by 7.6% and 7.1% respectively. Besides, the proposed algorithm performs better in stability, which can provide more accurate and personalized recommendation.