Optimizations for item-based Collaborative Filtering algorithm

Collaborative Filtering (CF) is widely used in the Internet for recommender systems to find items that fit users´ interest by exploring users´ opinion expressed on other items. However there are two challenges for CF algorithm, which are recommendation accuracy and data sparsity. In this paper, we try to address the accuracy problem with an approach of deviation adjustment in item-based CF. Its main idea is to add a constant value to every prediction on each user or each item to modify the uniform error between prediction and actual rating of one user or one item. Our deviation adjustment approach can be also used in other kinds of CF algorithms. For data sparsity, we improve similarity computation by filling some blank rating with a user´s average rating to help decrease the sparsity of data. We run experiments with our optimization of similarity computation and deviation adjustment by using MovieLens data set. The result shows these methods can generate better predication compared with the baseline CF algorithm.