In this paper,in order to solve the problem of data sparseness and improve the effect of recommendation,an improved collaborative filtering algorithm is put forward.Firstly,this algorithm calculates the item-types similarities through a new calculation method and the items whose similarities are greater than a certain threshold value will be considered as neighbors of the target-item.Secondly,the system predicts target-user's score values for the target-item according to the scores for the neighbors of target-item,and the predicted values will be filled in the sparse score matrix.Finally,this algorithm clusters the new matrix (K-means clustering) based on the users,to predict target-user's score values and make recommendations.The experimental results on the Movielens dataset show that this algorithm can effectively alleviate the data sparseness,reduce the computational complexity and improve recommendation accuracy.