In this paper, we propose a novel method combined classical collaborative filtering (CF) and bipartite network structure. Different from the classical CF, user similarity is viewed as personal recommendation power and during the recommendation process; it will be redistributed to different users. Furthermore, a free parameter is introduced to tune the contribution of the user to the user similarity. Numerical results demonstrate that decreasing the degree of user to some extent in method performs well in rank value and hamming distance. Furthermore, the correlation between degree and similarity is concerned to solved the drastically change of our method performance.