An improved collaborative filtering(CF) recommendation algorithm based on the second-order similarity between users was introduced.According to the empirical statistics,it was found that the definition of classical item-projection-based similarity contains so plenty of information of popular items that the users' interests or preferences are hard to be measured.Owing to considering the directive second-order similarity,the new algorithm can effectively depress the influence of mainstream preferences on target user.The numerical results on one benchmark dataset Movielens show that the accuracy can reach 0.080 8,which is improved by 22.08% comparing with the standard CF.Correspondingly,the diversity reaches 0.775 and could be improved by 10.87% in the optimal case when the recommendation list equals to 50.The study indicates that the directive second-order similarity are crucial in recommendation algorithm.