Research On Recommendation Algorithm Of Scientific Papers

With the development of science and technology, the number of scientific papers increases rapidly. At the same time, more and more researchers search scientific papers on the Internet to support the preliminary investigation and analysis. Currently, researchers retrieve papers through Google Scholar, CNKI and some other literature retrieval platforms, and many scientific papers recommendation platforms, which can solve the information overload problem, emerge as the time require.After researching and analyzing lots of recommendation algorithms, this paper put forward a new method of scientific papers recommendation based on language model through citation network.The three aspects of main works are as follows:Firstly, this paper introduces the critical technology of scientific papers recommendation, presents the concept and application of language model and word vector, analyses the value and application of citation network. Based on the similarity of citation sentence and the structure of nature language, the concept of the citation sentence is proposed for the first time. Each scientific paper is modeled as one word in the citation sentence, and then word vector can be used to describe scientific papers.Secondly, On the basis of citation sentence, this paper proposes the PaperLinkRank method and uses neural network language model tool word2vec to train and get the model. Simulation results on DBLP data set and CiteSeer data set show that the method works well, and performance indexes such as coverage, rank, and MMR are good. These results of the experiments show that the citation-network-based language model scientific papers recommendation system has a good performance. This method improves F1 measure a lot when compared with traditional collaborative filtering method. And compared with PageRank, the MMR is increased by about 15%, and the method has a better performance in terms of diversity.Thirdly, this paper designs and implements a recommendation module of a scientific information platform. It focuses on personalized recommendation process, implicit collection of user's behavioral data,data cleaning and analysis of the importance of actions to assign weight to data. At present, the system has collected more than 300 million scientific papers and has been developed as well as in the trial period.