Research On Personalized Recommendation Algorithms Based On Collaborative Filtering

With the swift and violent development of Internet, it’s increasingly difficult for users to acquire valuable information which meet individual preferences in the vast amounts of data. Personalized Recommendation System as an effective way to solve the information overload, has been widely used in e-commerce, social networking, advertising, film recommendation and other areas. Due to the advantages of simple and easy to implement, domain unrelated and highly accurate recommendation, Collaborative Filtering Recommendation has achieved great success and been wildly adopted in Recommendation System. It not only has great value in academic researches but also has broad applications in commercial. However, the inherent problems of data sparsity and cold-start in Recommendation System has been severely degraded the quality of recommendations. Therefore, it’s supported to make full use of the sparse data and incorporate related information. Most researchers used to adopting methods such as filling sparse rating matrix, changing method of similarity calculation, dimension reduction, incorporating trust in social network and so forth to solve the above problems.In response to the aforementioned problems, this paper focuses on Collaborative Filtering Recommendation algorithm based on matrix factorization method and variant which integrate the trust between users in social network. Recommendation based on matrix factorization can reduce the dimension of the original data to achieve highly accurate recommendation and have potential to cope with large-scale data. To further incorporate trust in social network into algorithms, we can get more information between users as well as resisting the interference of the malicious user in system. The major work is as follows:(1) We elaborate the meaning and present research status of Recommendation System. Furthermore we introduce different types of Recommendation System, including Content-Based Recommendation algorithms, Collaborative Filtering Recommendation and Hybrid Recommendation. Especially, we give a detailed description of collaborative filtering recommendation.(2) We introduce modeling and reasoning of Matrix-Factorization-Based recommendation algorithms and variant that incorporate trusts in social network. What’s more, we deeply explore the working principle and implementation mechanism of current typical model.(3) To improve the recommendation accuracy, we propose a novel algorithm called Trust Impact MF which integrates other users’ indirect influence on active user’s future ratings and trust between users in social network.(4) A new algorithm called Sim Trust MF is proposed based on the weighted relationship between users. The weighted relationship used by Sim Trust MF model is composed of reconstructed trusts between users and similarity between users. Reconstructed trusts between users is used to attach a great important to users who are trusted by many users in social network, while similarity between users is used to distinguish target user may have different preference to his or her trusted user.(5) A large number of experiments are conducted under different data sets in term of different situation and we analyze the results of the proposed algorithms with other classical algorithms. The results of experiments illustrate that the two proposed algorithms can both achieve better recommendation accuracy as well better handling cold-start problem.