一种融合用户显隐式阅读偏好的论文推荐模型  

A PAPER RECOMMENDATION MODEL WITH USER EXPLICIT AND IMPLICIT READING PREFERENCES

作　　者：唐浩 刘柏嵩[1] 黄伟明 Tang Hao;Liu Baisong;Huang Weiming(College of Information Science and Engineering,Ningbo University,Ningbo 315211,Zhejiang,China)

机构地区：[1]宁波大学信息科学与工程学院,浙江宁波315211

出　　处：《计算机应用与软件》2022年第5期253-259,共7页Computer Applications and Software

基　　金：学术性大数据知识组织与服务标准研究项目(15FTQ002)。

摘　　要：在海量学术文献的个性化推荐中,现存基于内容的方法以CNN作为特征提取工具,关注用户的显式阅读偏好,却忽略了全局语义特征,而基于图的方法通常忽略用户和论文之间的高阶关联结构信息。针对以上问题,提出一种混合推荐模型GNPR(Graph Neural Paper Recommendation),能够学习更完整的用户显式阅读偏好及用户和论文之间的高阶关联信息。该方法使用Word2vec和DCNN(Dual Convolutional Neural Network)处理文本,以双层自注意力的特征抽取模式学习文本全局特征,补充用户显式阅读偏好。针对概念、用户、论文和论文元数据等数据构建知识图谱,使用改进的图卷积网络学习用户和论文之间的高阶关联信息,从而挖掘用户隐式的阅读偏好。在CiteULike-a等数据集上验证了GNPR模型的有效性。In the personalized recommendation of massive academic literature,the existing content-based methods use CNN as a feature extraction tool,focusing on the user’s explicit reading preferences,but they ignore the global semantic features.The graph-based methods usually ignore the high-order association structure information between users and papers.Aiming at the above problem,this paper proposes a hybrid recommendation model GNPR,which can more effectively learn the user’s explicit reading preferences and the high-level association information between the user and the paper.This method used Word2 vec and DCNN to process the text,and learned the global features of the text with a two-layer self-attention feature extraction mode to supplement the user’s explicit reading preference.The knowledge map was constructed for the data such as concept,user,paper and paper metadata.The improved graph convolution network was used to learn the high-order correlation information between users and papers,so as to mine users’implicit reading preferences.This paper validates the effectiveness of the GNPR model on datasets such as CiteULike-a.

关 键 词：论文推荐 知识图谱 高阶结构信息 用户偏好 图神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]