A Study On Latent Friend Recommendation Based On Topic And Links On SNS

As a typical application of Web2.0era, Social Networking Site (SNS) quickly becomespopular all over the world, and has become an indispensable part of our daily life. Users onSNS build and extend their own social circles by adding and following other users as friends,and thus they can interact and share information with each other. However, with the rapidexpansion of SNS and the sharp increase of the numbers of users, it becomes too difficult forusers on SNS to look for friends and to extend their own social circles. In order to solve theproblem of information overload, recommender system on SNS emerges at the right moment.Users on SNS not only contact with users they already know in real life, but also want to addand follow users with similar interest, namely “Latent Friends”. However, there are fewstudies on latent friend recommendation, and no SNS website provides the application ofrecommending this type of friend, which greatly inhibits the development of SNS. In order tomeet the demand of making friends with similar interest, this paper focuses on latent friendrecommendation in MicroBlog for instantce.Currently, among all recommender systems on SNS, recommender system based on linkscan only recommend a very limited number of friends, while recommender system based oncontent has low accept rate and recognition rate. Therefore, this paper proposed a hydridrecommender model in order to improve the effect of recommendation. This paper introducesin Topic Model (TM) to solve the deficiencies of traditional vector space model (VSM). In therecommender module based on topic, the paper porposes a UserLDA model which is suitablefor SNS. The model gathers all the tweets with the same user, uses Collpased GibbsSamplling to estimate parameters, and converts user vector based on key words to that basedon topics. It helps to reveal users’ interst preference based on hidden topics, and thetopic-similarity between users can be calculated according to users’probability distribution ontopics. In the recommender module based on links, the paper improves RA index and Jaccardcoefficient in link prediction problem to make it applicable in a directed network, and thuscalculates link-similarity between users. Finally, the hydrid model calculates theomprehensive similarities based on a combination of topic-similarity and link-similarity,based on which recommends Top-N latent friends to a target user.In order to verify the effectiveness of the proposed model, this paper applies it to a realdata of sina weibo, and conducts a contrast experiment with VSM, and a two-phase friend recommender model. Experimental result shows that the proposed model has higher F1measure than the other two models, which means the proposed model can produce betterrecommendation.