A Network Graph Model For Scientific Article And Its Tag Recommendation Method

As tens of thousands of scientific articles being published every year,the scale of scientific articles reaches an unprecedented height.It has become a thorny problem for researchers to find their interested papers from the numerous scientific articles.Historically,researchers used the traditional collaborative filtering models to recommend scientific articles for users.The main problem of these models is that they can not solve the cold start problem,therefore,the content-based models were proposed.However,these models only exploit the textual content of articles and ignore the latent connection between academic papers,so the recommended results are unsatisfactory.In view of the shortcomings of these two models,this thesis focuses a network graph model to explore effective recommendation methods of new academic papers and the question of recommending suitable tags for academic papers.By analyzing the disadvantages of the network graph model and drawing on the advantages of easy integration of heterogeneous information,this thesis proposes two graph models: one integrating tag information for new scientific article recommendation and the other using multi-source heterogeneous graph model for tag recommendation.The former mainly integrates tags' information of scientific articles into their text content to compute the similarity between the articles,which effectively solves the problem of data sparseness caused by the lack of historical users' behavior information and enhances the reliability of intrinsic links between the content of articles.The latter expediently integrates heterogeneous information into the graph model for efficient tag recommendation and effectively solves the cold start problem in the tag recommendation.Apart from that,compared with other graph models,these two graph models incorporate only a few similarity of scientific articles,which greatly simplifies their structures,so they are well adapted to sparse data and require less computation cost.The models proposed in this thesis integrate the heterogeneous information.The random walk and restart algorithm is adopted to calculate the similarity between the nodes in these graph models,and then the scientific articles and their tags can be recommended.A series of experiments on two real datasets show that the two graph models which are proposed in this thesis not only significantly outperform other baseline models,but also require less running time.