Text Representation Learning Based Citation Recommendation Approach

The citation recommendation approach is mainly to automatically and effectively help researchers find a list of papers related to the researchers query information.With the publication of a large number of research papers,it is becoming more and more important for researchers to quickly and accurately recommend relevant references.Among the many citation recommendation methods,the graph-based citation recommendation method has attracted much attention because of its flexible integration of rich information,but it lacks the effective use of the content information of the paper.This paper based on text representation learning method,two graph-based methods are proposed to solve the problems of missing data information or lacking context information.The main work and achievements are as follows:(1)In order to solve the problem of missing data information,a citation recommendation method based on Deep Walk was designed and implemented.The recommended method mainly includes three processes.Firstly,a two-level graph model is constructed by using citation relationships,keyword information and the relationship between them,calculating the similarity of the content of the paper,according to the similarity of the papers,the adjacency matrix is regained,and the content information of the paper is integrated into the graph node;Then,a random walk algorithm with restart is used to obtain the context node,and a distributed representation of the nodes in the graph is generated using the Skip-gram model.Finally,the distance between the target paper and the candidate papers is calculated,and the N papers with the smallest distance are selected as the recommended papers for the target query.Compared with PW(Paper-Word),APW(Author-Paper-Word)and Node2 vec under the same data set and the same evaluation index,the citation recommendation method based on Deep Walk increased by an average of about 14% at recall@N,and the average increase in NDCG@N is about 10%.The higher prediction accuracy of this method is proved.(2)In order to make effective use of the content context information of the paper,a personalized citation recommendation method based on a three-layer graph model is designed and implemented.Through the three kinds of information of the paper,the author and the keyword and the mutual relationship between them,a three-level graph model is constructed,and different parameters are set in different types of attribute layers so as to achieve the purpose of distinguishing different types of attribute layers;Word2vec is used to generate the vector representation of the content of the paper.The similarity of the content of the paper is obtained by the cosine similarity calculation formula.The related information is used to generate the query vector and the score of the paper related to the query is improved.Adopting a Random Walk with Restarts(RWR)run recommendation model to generate the final recommendation results.The experimental analysis was performed in the same data set with the five reference methods of RTM(Relational Topic Model),Link-PLSA-LDA,Pop Rank,LDA(Latent Dirichlet Allocation),and Cite Rank.The experimental analysis was performed in the same data set with the five reference methods of RTM(Relational Topic Model),Link-PLSA-LDA,Pop Rank,LDA,and Cite Rank.The experimental results show that this approach is superior to the other baseline approaches and can achieve better recommendation results.An experimental comparison of the method in(1)and this method shows that the method in(1)is better than this method for both the recall@N and NDCG@N evaluation index,it proves the effectiveness of the feature vector representation obtained by the method in(1).