Research On Top-N Recommendation Problem Based On Social Network

With the rapid development of Internet, social networking services has becomemore and more popular in recent years, even become an important part of manypeople’s life. When bring people a lot of convenience, social networking services alsogive rise to the problem of information overload. Recommender system is one of theeffective means to solve the problem of information overload. However, many existingrecommender systems of social networking service are not good enough. Different withtraditional recommendation problems based on rating, recommendation problems ofsocial networking are mainly Top-N recommendations problems. We research on theTop-N recommendation problems of social networking illustrated by the example ofcelebrity friend recommendation of microblog and singer recommendation of onlinemusic system. The main contents of this paper are organized as follows.First, we research the celebrity friend recommendation of microblog as a linkprediction problem by the classical link prediction model, Logistic Regression Model.We use this model combining social network structure, keywords and item popularityfeatures. In these features, item popularity is better. Then, we research the celebrityfriend recommendation of microblog as a recommendation problem by the classicalrecommendation model, Nearest Neighbor Model. We use three kind of similaritymethods to compute similarity matrix. Then, we get a guideline of further study by theanalyzing the advantages and disadvantages of these two models.Secondly, we choose Matrix Factorization Model based on learning to rank byanalyzing the limitations of Logistic Regression Model and Nearest Neighbor Model tosolve Top-N recommendation problems of social networking. We use this modelcombining eight features which are social relationships, social action, user age andgender, keywords and tags, item categories, time dynamics, user action patterns anduser click intervals. The precision of our experiments results reachs the state of the artlevel.Finally, we propose a new method of expanding features by social relationships tosolve the problem of feature sparsity. We conduct experiments of five features on boththe problem of celebrity friend recommendation of microblog and the problem of singerrecommendation of online music system. The experiments results fully verify theeffectiveness of our method. Specially, the precision of our method on the problem ofsinger recommendation of online music system is better than the state of the art method.