Aiming at data sparseness and inaccurate user relationship in traditional collaborative filtering recommendation algorithm, an improvement recommendation algorithm based on asymmetric similarity relationship of users is proposed.By using the sample number of potential features and the decomposition of singular value matrix, the asymmetric similarity between users is calculated, and the relation between users is defined.The simulation results show that , the Mean Absolute Error of the algorithm is superior to the traditional algorithm with the increase of the number of neighbours,the minimal value is about 0.682 between the number of neighbors is 40-60. The value of Mean Absolute Error is about 0.758 of the traditional algorithm. It can be seen that the algorithm to determine the user relationship is more accurate, predictive score is relatively close to the actual score.
