Abstract:

Due to great advances in the mobile Internet, the Social Network Service (SNS) has become an indispensable service. The development of current mainstream social media shows a trend that social service and information service are combined and interwork to provide a better experience. Meanwhile there is increasing polarization among users. The heterogeneous features, say, the combination of information content and sociality as well as the polarization of user roles, present challenges to traditional research in social media. Some studies of social media are mainly based on the equal position among nodes or similar relations. If the algorithms brought about by these studies are applied directly to the networks where the users are highly polarized, the results may be distorted or even be quite different from the fact. A new model for interests mining based on social relations is proposed in this paper. Dealing with the polarization in social media, we incorporate matrix factorization and the label propagation algorithm to treat information disseminators and average users, respectively, in order to discover interests of average users in a large-scale heterogeneous network. The validness of the model and the performance and advantages of the algorithm are tested and verified in Zhihu datasets. Experiments show that the maximum increase in the recall of the proposed method, compared with the baseline, is 42%.

Key words: heterogeneous network, social networks, interest model, non-negative matrix factorization, label Propagation