Research On Citation Recommendation Based On Heterogeneous Information Network

In recent years,with the vigorous development of science and technology,scholars have taken writing articles as an important means to show their scientific research achievements,which has led to the rapid growth of the number of scientific research literature.Therefore,it is a challenging task to recommend references for scientific researchers accurately and efficiently.Citation recommendation aims at intelligently selecting suitable references for researchers from huge literature databases.Citation recommendation algorithm has strong practical value and broad application prospects.It can improve the efficiency of searching literature for researchers,enhance awareness of beginners in current research fields,and reduce the probability of errors and omissions to a certain extent.The current research object is heterogeneous information network.However,most of the networks are heterogeneous in the real world,such as the computer academic literature network derived from DBLP.In the problem of citation recommendation,we can use a variety of information between articles to solve the problem,but this information is heterogeneous.So,using heterogeneous infonnation network to model citation recommendation problem can better capture potential relationship between articles,including semantic information and structural information.Current citation recommendation algorithms are based on the semantic similarity of articles,including author cooperation infonnation,article publisher information and so on.They use graph to model the cooperation information of authors and publishers,and use the topological structure of graphs to calculate the similarity between articles to generate recommendation lists.Different from graph model,heterogeneous information network includes not only the internal relationship of features,but also the relationship between features.In this dissertation,semantic information and structural infonnation are fused through heterogeneous information networks,meta-structure which under different weights is introduced to calculate similarity between articles.A new similarity calculation index is proposed to generate recommendation lists.Then,in order to solve the problem that meta-structure cannot measure the similarity between points without path connection in heterogeneous networks,representation learning is applied to solve the citation recoimmendation problem,so that the vectors of articles can learn the potential relationship between articles.Lastly,our algorithms calculate similarity and recommend citations according to similarity ranking.The main work of this dissertation is as follows:(1)Firstly,this dissertation elaborates the research significance of citation recommendation problem and the research status at local and abroad,focusing on the advantages and problems of solving citation recommendation problem based on heterogeneous information network.Then the dissertation briefly describes the development of heterogeneous information networks and citation recommendation problems,and gives the significance of introducing heterogeneous information networks into citation recommendation problems.(2)To solve the problem that multiple features cannot be captured simultaneously to calculate the similarity between articles,this dissertation proposes a meta-structure-based citation recommendation algorithm(MS-AIOA).Firstly,the meta-structure within three degrees is selected through the three-degree influence principle.Secondly,different similarities are calculated based on different meta-structures,and a new similarity calculation method combining multiple meta-structures is proposed to balance the weight of each meta-structure.Finally,an intelligent optimization algorithm is introduced to solve the weight parameters of meta-structure,which can achieve the best recommendation effect.Experiments show that the algorithm improves the recall,accuracy,F1 value and NDCG of citation recommendation.(3)The meta structure-based citation recommendation algorithm cannot measure the similarity of two articles without reachable paths in the network.To solve this problem,this dissertation proposes a citation tendency-based citation recommendation algorithm(CRW).This algorithm is divided into two parts:Firstly,obtaining the walk sequence of nodes in the network by random walk based on Citation tendency;Secondly,sending the walk sequence into Skip-gram to learn the relationship between the article nodes,so as to obtain the vector representation of the article nodes.In the process of random walk,the algorithm gives five walk weight matrices,which correspond to five types of edges in heterogeneous information networks.Among them,the value of weight matrixes make the direction of random walk more inclined to the citation of articles in the sequence.Finally,giving the method of calculating similarity and generating recommendation list for two article vectors.Experiments show that this method is superior to the meta structure-based citation recommendation algorithm(MS-AIOA).