To address the problem that the traditional collaborative filtering algorithms only use user's rating data to compute the user similarity, which leads to a poor recommendation precision, an improved collaborative filtering recommendation algorithm was put forward. Firstly, the user's score difference level was obtained by using user's average score as the boundary point, which was considered as a weighting factor in the user's similarity. Secondly, according to the user's rating data and the item category information, the user's interest level for the item category and the users item preference were mined to calculate the user's preference similarity. Thirdly, the above two similarities were combined to get the intergrated similarity between users. Finally, the traditional item similarity and the intergrated similarity between users were fusioned to predict score and recommend items. The experimental results show that, compared with the traditional user-based collaborative filtering recommendation algorithm, the Mean Absolute Error (MAE) of the proposed algorithm is reduced by 2.4% on average. The new algorithm can effectively improve the accuracy and quality of the recommendation algorithm.