Research On Data Sparsity Problem In Collaborative Filtering Systems

As an effective solution to information overload, recommendation system has been widely used in electronic commerce, social network,. online music and movies communities,etc. Because of its easy implementation,low data dependency, and accurate results, collaborative filtering recommendation algorithm has become one of the most widely used technologies in the recommendation system. Collaborative filtering recommendation system mainly includes two methods, which are based on memory and model respectively. However, these algorithms have the problems, such as data sparsity, cold start and system scalability. In this paper, aiming to help alleviate the problem of data sparsity, the research work:(1) advanced a collaborative filtering recommendation algorithm based on candidate items set (CI-CF). In this algorithm, the asymmetric impact and the support level between users were firstly incorporated. And then a new concept of candidate items set has been defined to measure users’preferences on the items. Moreover, this algorithm use the item entropy to retune the preliminary items which haven’t been rated. Experimental results on Movielens dataset and Netflix dataset demonstrated that the proposed algorithm outperformed many state-of-art algorithms(AC-PCC, RA-CF, User-CF) in measurements of accuracy, recall and F1 values.(2) proposed a hybrid collaborative filtering recommendation algorithm based on friendships and tag (FT-CF). The algorithm alleviated the data sparsity with the help of social network relationships and social tag information. Firstly, it utilized Adamic-Adar Index to analyze the friendship networks, found out the candidates similar to a target user, and based on them, ascertained the commodities which satisfied the target user’s demands. And then, in order to further ameliorate the data sparsity, the algorithm introduced the idea of TF-IDF to analyze the user’s interests from the historical record tags. Finally, the algorithm combined information from both social network relationships and historical record tags effectively. Experimental results on Lastfm dataset exhibited that the proposed algorithm outperformed many state-of-art algorithms (PRT-CF and UCTRA) in measurements of accuracy and recall.(3) Summarized varieties of algorithms based on different data information, such as the tag information data,the grading data and user/item properties. The Personalized Resource Recommendation Based on Tags and Collaborative Filtering, the Biclustering neighborhood-based collaborative filtering method for top-n recommender systems and Coupled Object Similarity Based Item Recommendation Algorithm were introduced, elaborated, and compared comprehensively. Experimental results on MovielenslOm dataset showed that the more information from users/items, the more understanding on them and the more easily grasping the users’ interests. Therefore the effective combinations of a variety of data information are helpful to improve the quality of the recommendation system.