Research On Meta-path Based Citation Recommendation

With the popularization of the Internet and the continuous development of technology,scholars publish their results in the form of papers or patents.The advancement of science and technology has promoted the development of the times,and at the same time,the number of papers or patents has increased exponentially.The massive scientific and technological literature provides sufficient support for the research of scholars,but an ensuing problem is how to choose suitable literature for the study.Therefore,how to efficiently and accurately provide scholars with suitable literature is an urgent need to solve the problem.A citation recommendation algorithm is an effective method to solve this problem.Citation recommendation aims to automatically recommend appropriate references for given academic literature,and it has received extensive attention from scholars in recent years.In the citation recommendation question,the citation is not only related to the content of the article,but also closely related to the author of the article and the relevant institution of the article.The citation network is modeled as a heterogeneous information network,and the rich semantic information in the citation network can be mined through meta-paths.The current graph-based citation recommendation algorithm mainly regards the citation recommendation task as a link prediction problem,uses graph models to model the citation network,uses network representation learning to obtain the potential representation of the article,and recommends the corresponding article by calculating the similarity between the articles reference list.The difference from the pure graph topology is that the heterogeneous information network not only includes the graph topology but also covers the attribute information of various objects in the network.This thesis uses a heterogeneous information network to model it,integrates text attributes and topological structure,combines meta-path and random walk to obtain the sequence representation of the node,uses the network representation to learn to obtain the node vector representation,and recommends according to the similarity ranking.Since a single meta-path can only represent specific structural semantics,in order to reflect the complex relationship between objects in a heterogeneous information network,multiple meta-path fusion is used to obtain the characteristics of the node,and the heterogeneous network representation learning is used to obtain the potential representation of the node.The multi-element path is a collection of many different types of meta-paths.The multi-element path contains more structural semantics and provides more structural features in the subsequent representation of learning nodes.The main contributions of this dissertation are as follows::(1)This thesis proposes a Citation Recommendation algorithm based on Meta-path Hybrid Random Walk(CR-MHRW)to solve the citation recommendation problem.This thesis constructs semantic links based on text attribute similarity and keyword semantic similarity,and fuses the semantic similarity in the article into the constructed heterogeneous information network.To obtain the latent representation of the node,each node is sampled using a combination of meta-path and random walk,and the network representation learning model is used to learn the latent representation of the article node,and finally generated by calculating the similarity between the article nodes recommended list.Experimental results show that the proposed method can effectively integrate article node attribute information and topological structure.(2)For a single meta-path,it is difficult to reflect the complex relationship between multiple objects in the citation network.This thesis proposes a Citation Recommendation algorithm based on Multiple meta-Paths Fusion(CR-MPF).This algorithm mainly consists of three parts.First,build a heterogeneous information network that includes citation relationships based on text attribute similarity and keyword semantic similarity;then through the fusion of multiple meta-paths,meta-paths of multiple semantic structures are combined with random walks.Combining to obtain a variety of potential relationships between articles;Finally,the potential relationship between articles in the citation network is obtained through the heterogeneous information network representation learning model,and the similarity between article nodes is calculated to generate a recommendation list.The experimental results show that the fusion of multiple meta-paths can effectively capture the potential relationships between articles in the citation network.