Research On Social User Model And Its Applications In Recommendation Systems

As a communication platform, social network can be used to publish, obtain, spread, and comment the messages easily. Microblogging is a kind of popular broadcast social media, which is based on the mechanism of follows. The real-time updating of microblogging in social networks can produce large amounts of unstructured data, which leads to the challenge of information overload and information redundancy. Therefore, people will spend more time in filtering out and selecting relevant information. How to help users to get the contents they desire from vast amounts of information is one of hot issues to be solved currently.Based on analyzing the historical data of users, the user modeling technology extracts the interest subjects, which has already been applied to many different fields, such as e-commerce, advertising and marketing, search engine and personalized recommendation systems. As a form of short text, microblog content presents the characteristics of fragments, which poses challenges for traditional user models in the aspect of obtaining their interest subjects with the long text. Additionally, there exists other complex relations between users except for the relations of interest, such as relations of follows, comments and forward, which are difficult to be described by the traditional user model.In this paper, the proposed social user model not only can describe microblog users’ interests, but also can depict the relations among them. Based on social user model, we have explored deep researches on neighborhood user model, multi-granularity similarity relations of subject, diversity of hot topics and evolution of community interest, etc. With the massive experimental data gathering from the microblogging platform, the effectiveness of these methods have been verified, which can be used to solve the problems of diversifying users’ interests, improving prediction accuracy and diversifying recommendation results in personalized recommendation systems. The specific research results are as follows:(1) The social user model based on users’ relations is proposed. In the microblog scenario, the accuracy of interest is influenced by the fragments of short text. In this paper, we fully take into account social interaction relations among users, and fuse these relations into the process of obtaining their interests. Furthermore, we have built the Social User Model Based on Users’ Relations(SUM-UR), which can be taken as the basis for further study, such as expansion of user’s interest subjects, discovering of hot topics and evolution of community’s interests.(2) Interest subject expansion based on neighborhood user model is presented. In order to solve the sparsity problem of interest subjects for short text in the microblog scenario, we propose the concept of perception relationship to define the neighborhood of a user, which is considered as the combination of resource perception relationship and follow perception relationship. Based on the neighborhood set, we update the set of interest subject and interest degree of subject for a user, which implements Interest Subject Expansion Based on Neighborhood User Model(ISE-NUM). Experiments show that the neighborhood user model based on perception relationship can effectively expand the set of user’s interest subjects. Comparing against the collaborative filtering recommendation method, the proposed method can have higher precision and recall rate with a rise at about 10% respectively.(3) Friend recommendation based on multi-granularity subject similarity relationship is investigated. The recall rate is low in the aspect of friend recommendation at one subject. In this paper, we propose a method on how to recommend friends in terms of a multi-granularity hierarchical subject similarity relationship. Firstly, we calculate the multi-granularity level of similarity by considering the content interest of subject and structure of semantic interest tree, and then make a study of multi-granularity similarity relationship between users’ models. Meanwhile, we implement recommending friends by the mechanism of multi-granularity subject similarity relationship. Experiments show that Friends Recommendation Based on Multi-granularity Subject Similarity Relationship(FR-MSSR) is superior to the traditional Fo F friend recommendation approach in the aspect of precision and recall.(4) Diversified microblog recommendation based on personalized hot topics is researched. A burst of a hot topic can be induced by the microblogs with large numbers of comments and forwards. The similarity of contents in the traditional recommendation methods is much too large, which leads to the poor diversity of results. In this paper, we put forward a method on how to calculate the popularity degree of subject, and predict the popularity degree with smooth exponential curve. Furthermore, the hot topic is discovered. By combining user’s interest, we investigate Diversified Microblog Recommendation Based on Personalized Hot Subject(DMR-PHS), and propose a Max Min distance set selection algorithm in order to find top-k diversified microblogs. The experimental results show that the prediction results of popularity degree are accurate and the MAE is about 0.07. Meanwhile, the satisfaction degree of users can be promoted by diversified microblog recommendation of hot topic.(5) Community interest evolution based on key users is analyzed. The dimension of community interest is high in the microblog scenario, which makes the evolution analysis of community interest more difficult. This paper proposes the community interest trend analysis method in terms of representative roles of key users. The method is at first to dig key users from the user’s community, and then analyze the changes of key users in different time windows in order to reflect migration of community structure. Then, Community Interest Evolution Based on Key Users(CIE-KU) is proposed. Experiments show that the method can not only effectively cover interests of community, but also can well simulate the evolution process of community interest.