Research On Citation Recommendation Method Based On Multi-Information Integration And Heterogeneous Hypergraph Neural Network

Driven by the technological revolution of the Internet,a vast amount of academic information is growing rapidly online,making it difficult for researchers to retrieve the most appropriate literature efficiently and accurately from databases.Citation recommendation aims to use the metadata information embedded in the target papers to design models that help researchers quickly and efficiently find relevant citations that match their information needs and research interests.Most existing local citation recommendation methods make use of contextual textual data to learn semantic representations,yet some auxiliary information unrelated to the textual content is often also closely linked to researchers’ citation intentions.Therefore,modeling accurately and effectively is extremely important for achieving accurate citation recommendation tasks.In addition,in the graph-based global citation recommendation methods,the complexity of node types and interactions between nodes can be exploited to make the most of the diversity and plurality of this information helps to understand the real needs of researchers.By summarizing and analyzing existing citation recommendation models and methods,this paper investigates citation recommendation methods based on multi-information fusion and heterogeneous hypergraph neural networks,which mainly include:(1)To solve the problems of the limited scope of information consideration on the auxiliary side and too coarse information granularity in existing local citation recommend ation methods,we propose a local citation recommendation model based on a dual attention mechanism and multi-information integration.The method is divided into two modules.The first module contains three information encoders,which are mainly responsible for learning the embedding of the target paper.Specifically,the contextual information encoder learns the textual features of the citation context using the dual-attention mechanism,the historical citation information encoder models the semantic information provided by the cited literature using the text embedding network,and the author information encoder fully explores the user’s research area,interest preferences and their implicit information from users’ perspective.The three types of information are aggregated through a weighting strategy,taking into account their different levels of contribution to the literature recommendation task.The second module is responsible for vector learning of candidate papers and calculating the correlation between candidate papers and target papers to achieve local citation recommendation.(2)To address the problem that the existing global citation recommendation methods have insufficient multiplicity and diversity representation of node types and interaction relations,we proposed a dual channel heterogeneous hypergraph neural network for global citation recommendation.Specifically,we first construct a heterogeneous graph with the vertices of papers,authors,keywords,and venues,encode the local and global semantic features of each node in the heterogeneous graph using convolutional neural network and Transformer respectively,to obtain the structural representation of the target node in the heterogeneous graph channel.Secondly,taking the advantage that hypergraphs can flexibly model complex relationships,hypergraphs were created based on heterogeneous graphs and three different types of hyperedges were designed to extend heterogeneous data information.The interaction between nodes is encoded using the hypergraph structure and the potential complex higher-order semantic relationships in the hypergraph are captured using a hypergraph neural network to obtain a semantic representation about the target node on the hypergraph channel.Again,the information on the above two channels is organically combined to obtain the final semantic representation of the target node.Finally,relevance matching is performed with the candidate paper nodes to generate a citation recommendation list.(3)Finally,extensive experiments have been conducted on several publicly available datasets,and the results show that the method proposed in this paper outperforms existing citation recommendation methods,demonstrating the effectiveness of the proposed method.