In most traditional collaborative filtering algorithms, user rating matrix is usually an only important factor in calculating users’ similarity instead of considering the impact of the correlation between items. Therefore, an improved model of collaborative filtering recommendation is presented in this paper. Firstly, a method of item similarity measurement is introduced in the process of computing the user’s nearest neighbor to get more appropriate neighbors. In addition, due to that the users interests will decay over time, time weight is added in the process of computing item ratings．Experimental results show that the proposed algorithm can obtain better performance than other traditional collaborative filtering algorithms in aspects of prediction accuracy and classification accuracy.