With the development of personalized recommendation systems, the research of collaborative filtering reached a bottleneck. Neither algorithm accuracy nor computational complexity can be improved significantly. In this paper, we present our statistics and analysis on some recognized datasets. The analysis shows that the real rating features of the users cannot follow even distribution while most current algorithms were based on this premise. Therefore we proposed a new user-based collaborative filtering algorithm combining data-distribution. Since different users have different rating ranges, the key method of the algorithm is the special revise of user preference according to the distribution of the ratings. Our algorithm is comparable in computational complexity to SLOPE ONE algorithm and more accurate on the sparse data. We believe that it is a hopeful new direction for the development of collaborative filtering, which reflects the highlight of this paper.