Study On The Sparsity And Accuracy Problem In Collaborative Filtering Recommendation Technology

Collaborative filtering recommendation technology is one of the most widely used personalized recommendation technologies, whose core idea is that through calculating the similarity between users (or items) by using the existing user-item rating information, a/an user (or item) finds the set of users (or items), whose interest (or behavior) is similar to the user (or item), and then the recommendation system makes assessments, predictions and recommendations based on the rating value of the set of users (or items). Due to the cold start problem caused by new users or new items, and the user’s rating habit, that most users are accustomed to giving only a small amount of ratings or no ratings directly, the proportion of rated items becomes less and less, and the user-item rating information table becomes sparser and sparser, as the number of users and items increases rapidly. In addition, the traditional item based collaborative filtering recommendation technology calculates the similarity between one item and another basically according to the user-item rating information table, while ignoring the impact of the similarity between one item’s properties and another, and the time interval between two rating activities. Under such circumstances, the recommendation accuracy of the collaborative filtering recommendation technology can be affected to a certain extent, and can not achieve the expected results.For this reason, this paper makes an in-depth analysis about the sparsity and accuracy problems emerged in the actual recommendation system’s collaborative filtering recommendation technology, and proposes appropriate solutions. Details are as follows:1、In order to alleviate the data sparsity problem of collaborative filtering recommendation technology, an improved algorithm, based on the existing algorithm, is introduced to further alleviate the sparsity of data set and cold start problem and to improve the recommendation accuracy. The algorithm determines the current user’s user group, analyzes the general score of the user group on different item attributes, synthesizing the rated data information, and then fills the user-item rating information table, thus alleviating the data sparsity to some extent. Finally, the algorithm predicts ungraded items and generates recommendation lists. Simulation results show that the improved algorithm gains better prediction accuracy and further reduces the sparsity of data set and cold start problem.2, In order to improve the accuracy of similarity in collaborative filtering recommendation technology under special circumstances and the accuracy of collaborative filtering recommendation technology, this paper proposes two improved measurement algorithms:①a collaborative filtering recommendation algorithm with time adjusting based on attribute center of gravity model, on the basis of the improved collaborative filtering recommendation algorithm based on user characteristics and item attributes, and②a collaborative filtering recommendation algorithm with time adjusting based on Vague set, on the basis of the improved collaborative filtering recommendation algorithm based on user characteristics and item attributes. The first algorithm uses the filled user-item rating information table to determine each item attribute’s weight and uses the attribute center of gravity model to calculate the preliminary similarity, then makes adjustment to the composite similarity of preliminary similarity and traditional similarity to get the final similarity. The second algorithm uses Vague set theory to get Vague value from user-item rating information, and calculates the similarity, between one Vague value and another, as the preliminary similarity, and then makes adjustment to the composite similarity of preliminary similarity and traditional similarity to get the final similarity. Simulation results show that the two algorithm gain a higher recommendation accuracy, compared with the improved collaborative filtering recommendation algorithm based on user characteristics and item attributes.