Research On Social Recommendation Based On Matrix Factorization

The recommendation system,which is a data mining technology based on interactive rating data,extracts personalized preferences of users and items by means of deeply mining the potential information from explicit and implicit feedback data and predicts the degree that a user is interested in an item.However,there exist some bottlenecks in practical applications of the recommendation technology,such as cold start of data,data sparsity and dynamic incremental data.As is well-known,data sparsity is one of the most common problems in the field of the recommendation system.The recommendation aims to mine the user's preference for the item from the explicit information and makes reasonable recommendations for users.However,the model cannot learn the user's interest to the maximum because of the lack of rating information,which leads to a large deviation between the prediction and the true and reduces the accuracy of the recommendation.For the problem of data sparsity,the thesis has carried out a series of the researches.The main work of the thesis is as follows:1.To alleviate data sparsity,the thesis introduces implicit feedback information for the recommendation model,such as social information,in addition to data mining of explicit data.Due to sharing the interest of social users,the model is able to deeply learn the user's interest preferences.In the thesis,matrix factorization is used to project the explicit information and social information to the corresponding low-dimensional feature space respectively and make the user's preferences and personalized feature of the item that are used to recommend items for users.2.For the accuracy of the model,the thesis proposes multiple recommendation strategies on the foundation of explicit and implicit information,including matrix factorization for recommendation,user-based recommendation and item-based recommendation.The recommendation strategies make recommendations from the perspective of potential factors of users and items,the user's social factors and characteristic factors of items.While using matrix factorization to decompose explicit and implicit information,the thesis combines the recommendation strategies into a predictive scoring mechanism to jointly learn the user's preferences and the characteristics of items.The method is helpful to improve the prediction accuracy of the recommendation model.3.So as to fully mine the social network,the thesis learns the personalized preferences of social users based on the user's behavior and uses feature vectors to represent them.Then,the difference of the social user's interests is used to measure the degree of social relationship between users,which represents the similarity of interests between social users.Finally,the fine-grained social information is embedded into the matrix factorization as implicit information to further learn the user's personalized preferences and improve the accuracy of the model.