Font Size: a A A

Research On Matrix Factorization Recommender System With User Relationship

Posted on:2017-01-16Degree:MasterType:Thesis
Country:ChinaCandidate:Z Q ZhangFull Text:PDF
GTID:2308330482487162Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
In the Internet era of the information with immeasurable speed of growth, it becomes more and more difficult for users to search for useful information effectively. In order to solve this problem, recommender systems emerge to meet different needs. A recommender system as a new computing model, tries to filter the information which the target user may be interested in. Most of the existing recommendation algorithms are based on the thought of collaborative filtering, a concept coming from the natural law of behavior analysis for people choosing goods. Typically, users always first consult experienced people for unknown items, and then make a decision according to the experience of these people. Collaborative filtering algorithms apply this idea on the recommender systems. However, collaborative filtering has three main problems:the data sparsity problem, cold start problem, and scalability problem. To alleviate these problems, social-based recommender systems were proposed. A social network is formed by the Internet interaction among users, which reflects the relationships of people in the real world. As a result, combining social networks and recommender systems will form a social-based recommendation system which can enrich recommendation results.Matrix factorization (MF) model is a classic model in recommender systems for its good expandability and social information is easy to be integrated into the model. Under this mechanism, extracting a group of desired similarity is a key factor to improve the accuracy of the recommender system. Traditional similarity functions judge user’s similarity on their feedback for the same item without looking at their in social relations. This can produce misleading results because a user’s feedback may not reflect his true interest. To address this problem, this thesis makes an in-depth analysis in the degree of users’relationships. We propose two new similarity metrics on two functions to improve the measure of users’ similarity in interest. To use them in the MF-based recommender system, we first present two kinds of social regularization based on the two similarity metrics respectively.Our further analysis shows that one new similarity metric reflects a one-to-one interest similarity between two users, and the other reflects the average similarity between target user and his/her friends. Besides, the two similarity metrics are based on the density of the social network. If the social network is very dense, the value of these two metrics approximate to user’s interest similarity. In order to mine the values of a social network, we integrate these two metrics into one social regularization and discuss their combination in different ways. Compared with the previous social regularization model, our proposed approach can explain the user interest similarity more clearly and decrease the sparsity in social networks. Finally, the experimental analysis shows our approach performs better than the previous methods in accuracy.
Keywords/Search Tags:Recommender System, Social Network, Matrix Factorization, Social Regularization
PDF Full Text Request
Related items