Research On Scientific Paper Recommendation Based On Graph Neural Network

Scientific paper recommendation systems aim to help researchers quickly and accurately find the high-quality academic papers they need from the vast database of online academic resources.After decades of development,especially the explosion of deep learning technology in recent years,scientific paper recommendation systems have made great progress.Nevertheless,existing scientific paper recommendation systems still have the following two shortcomings.First,they cannot effectively integrate semantic information and high-order interaction information.The existing scientific paper recommendation methods not only ignore the high-order interaction information between entities,but also can only mechanically connect semantic information and interaction information in most cases,lacking a unified organic framework to integrate these two kinds of information.Second,there is a lack of deep representation learning research on academic entities such as users and scientific papers.Entity representation learning has positive significance for improving recommendation quality,while existing scientific paper recommendation systems basically use representation learning methods from other fields directly,which is not very helpful for the improvement of the quality of scientific paper recommendations.To address the above two shortcomings,this study accomplishes the following two research works by mining academic information networks based on graph neural network technology:Firstly,the semantic integration idea is combined with graph neural network,and the pre-trained natural language model is fine-tuned by using citation network to generate the representation vector of the paper and fuse the semantic information.Then,using graph neural networks,high-order connectivity information in the user-paper interaction network is captured to further optimize the representation vectors of user and paper nodes,which are finally used for academic paper recommendation.In this section,this study introduces graph neural networks to the scientific paper recommendation task,which effectively mines the high-order interaction information between users and papers while preserving the semantic information of academic papers.Subsequently,this paper does further in-depth research for the entity representation problem in graph neural network-based recommendation systems.In this section,this study first generates scientific paper representations based on the topic document representation approach,and then trains the representation vector by aggregating topic information and interaction information between papers in the citation network using a graph neural network-based self-encoder.Based on the same paradigm,user interest vectors are generated using users’ paper interaction lists,and then user node representation learning is tackled by aggregating user interest information and inter-user interaction information in the user network.The experiments demonstrate that deep academic entity representation is positive for the enhancement of recommendation systems.