Node-user Based Matrix Factorization Model For Recommendation Algorithm

Matrix factorization(MF)based recommendation algorithms have been widely used and developed.Meanwhile,with the fast development of social network,users' social relationship plays a more important role in recommendation algorithms.The biggest difference between traditional social relationship and online social relationship is that in social network,information is significant influenced by some users with huge influence,who's called the “node users”.Traditional MF based models and social based MF models are proposed recent years,however there's few research on the influence of node users and how it affects the recommendation results.The contribution of this thesis consists of three parts: a)Analysis the factor of user influence in social network and define node users,with Spearman's rank correlation coefficient to determine the correlation of different factors;b)Combine user influence with MF model,propose node user based MF model;c)Propose two methods to extend social relationship for social network relationship data to solve the data sparsity problem,thus to improve the prediction accuracy.First,we analysed user influence in social network and defined node user,proposed node user based MF model and did experiments on both Douban Movie dataset and Yelp dataset.The results shows that this model works well and can reduce the root mean square error(RMSE)when the data quality is good enough.The results also shows that the factor of user influence differs in different circumstances.Second,we applied the two social relation extend methods and the results show in some cases,these two methods can lower the RMSE even further along with the node user based MF model.