User-based Collaborative Filtering (CF) algorithm offers recommendations to users by analyzing the preferences of similar uses. The Key step of this algorithm is to calculate the similarity between users based on the user-item rating matrix. The Pearson Correlation Coefficient (PCC) is the commonly-used measurement. However, when the ratings are sparse or unbalanced, it cannot represent the similar relationship accurately. This paper investigates the calculation of the similarity among users by adjusting the positive and negative similarity and transferring the similar relationship in social network. The experimental results on the extremely sparse data show that the proposed method can enhance the prediction and recommendation accuracy than the original method.