In this paper, as an algorithm used for collaborative filtering, there are some shortcomings about slope one algorithm in commercial recommendation system, such as the rating predictions without considering the behaviours and attributes of the users and item, and data sparsity. We proposed the improved slope one algorithm based on the singular value decomposition technique and item similarity to improve the algorithm and process. Then the implementation scheme and flow chart of the improved algorithm is given. Finally, the new algorithm is evaluated by four different datasets. The result shows that in sparse datasets the improved slope one algorithm is more precise than slope one algorithm. In addition, in the four datasets with different sparsity degree, the improved slope one algorithm is stable, and the change of MAE value is relatively stable.