Nowadays, the collaborative filtering techniques have demonstrated an excellent performance in the top-N recommendation. However conventional methods in similarity measurement are insufficient when the condition of data sparsity and cold start occur, which leads to a poor accuracy in prediction. In order to concur the limitation, a collaborative filtering algorithm of calculating similarity based on item rating and attributes is proposed. Firstly, we calculate the similarity of item attributes, then calculate the similarity of the project according to the user rating of the project. Meanwhile, a weighted control coefficient is proposed to combine the similarity between item attributes and rating of items, which contribute to obtain nearest neighbors. Experiments have shown that our algorithm has major potential in solving the problem of cold start, therefore improving the precision of the recommendation system.

When the traditional collaborative filtering algorithm faces high sparse data, its precision and quality of recommendation become unsatisfied. With the development of social networks, it is possible to selectively fill the missing value in the user-item matrix by using the friendship or trust relationship information of social networks. According to the memory-based collaborative filtering algorithm, in the paper, the two steps which are similarity calculation and user rating prediction are taken into account. Besides, this paper has filled appropriately the missing value and improved memory-based collaborative filtering recommendation algorithms to integrate the social relations. The experiment on the Epinions dataset shows that the improved algorithm can effectively alleviate the sparsity problem of user rating data and perform better than other classic algorithms in RMSE and MAP evaluation metrics.