Microblog User Recommendation Based On Trust Relationship And Topic Analysis

The rapid development of social network sites deeply impacts how people share information and communicate. As open online social network platforms, the success of social network sites depends on the degree of user interaction and stickiness. With the explosive growth of user scale, recommending potential friends accurately plays an increasingly vital role for more social networking site operators.This thesis discusses several friend recommendation mechanism currently used. Considering that traditional recommendation algorithms ignore the issue of trust in recommendation process, we design a social network trust model to measure user reputation and trust relationship between users. Then we filter users whose credibility is less than a threshold, and divide the trusted user into subgroups. Lastly, we generate the final recommendation list based on topic similarity. This paper mainly focused on helping people expand their circle of friends and improve user stickiness. In this thesis, the main work is as follows:Firstly, we take some information into consideration to model the trust relationship between users. These information includes graph structure of social networks, interact strength between friends, and topic similarity between users. Based of these information, we construct a heterogeneous graph which models the social network, users interaction and topic information in a unified framework for considering the strength of trust relationship jointly. In particular, we present an optimization formulation which is generated from the definition of trust relationship to compute weights of the three types of information.Secondly, we use the topic model on the whole document posted by a user and grouped by hashtag to get the semantic information instead of using microblogs one by one because microblogs are short, unstructed and noisy and not suitable for semantic analysis. With the enhancement of the content in a document, the number of cooperation between terms are risen and the topic model can be more suitable.Thirdly, the creditworthiness is introduced into the model for untrusted users’ detection. To solve the problem of insufficient labeled samples in large-scale social networks, we compute the creditworthiness of each user from a small seed set. We use the trust-propagation model posted in the previous step to update the creditworthiness of each user and remove users who have creditworthiness less than zero as untrusted users instead of machine learning models.Finally, we propose a community-detection algorithm for massive-scale social networks based on the trust relation model. In the algorithm, we utilize a new type of trust, the recommend trust which is inspired by PGP, to compute the trust degree between users iteratively. In particular, we use the insight of MapReduce, a parallel programming framework, to modify the algorithm and make it more scalable.