| 基于无监督图神经网络的学术文献表示学习研究 |
| |
| 引用本文： | 丁恒,任卫强,曹高辉.基于无监督图神经网络的学术文献表示学习研究[J].情报学报,2022,41(1):62-72. |
| |
| 作者姓名： | 丁恒 任卫强 曹高辉 |
| |
| 作者单位： | 华中师范大学信息管理学院,武汉 430079 |
| |
| 基金项目： | 国家自然科学基金青年科学基金项目“基于深度语义表示和多文档摘要的学术文献自动综述研究”(71904058);中国博士后科学基金项目“面向综述自动撰写的摘要式学术搜索引擎研究”(2020M682458)。 |
| |
| 摘 要： | 学术文献特征表示,是学术文献搜索、分类组织、个性化推荐等学术大数据服务的关键步骤。研究表明,图神经网络能够有效学习文献的特征表示,然而当前研究主要集中在有监督学习方法上,不仅对数据集的大小和质量的要求较高,且学习到的文献特征表示与具体任务高度耦合。基于此,本文将四种无监督图神经网络方法引入学术文献表示学习,从Cora、CiteSeer和DBLP (database systems and logic programming)数据集的引文网络、共被引网络和文献耦合网络中学习文献的表示向量,并应用于文献分类和论文推荐两大下游任务。研究结果表明,(1)深度互信息图神经网络适合于文献分类任务,对抗正则化变分图自编码器则在论文推荐任务上性能更佳;(2)Cora数据集上的结果表明,相较于共被引和文献耦合网络,引文网络更适合于学习通用的文献表示向量。
|
| 关 键 词： | 无监督学习 图神经网络 表示学习 学术文献 |
Using Unsupervised Graphs of Neural Networks for Constructing Learning Representations of Academic Papers |
| |
| Ding Heng,Ren Weiqiang,Cao Gaohui.Using Unsupervised Graphs of Neural Networks for Constructing Learning Representations of Academic Papers[J].Journal of the China Society for Scientific andTechnical Information,2022,41(1):62-72. |
| |
| Authors: | Ding Heng Ren Weiqiang Cao Gaohui |
| |
| Institution: | (School of Information Management,Central China Normal University,Wuhan 430079) |
| |
| Abstract: | Constructing feature representations of academic papers is a key step in providing scholarly big data services,such as academic searches,literature classification and organization,and personalized paper recommendations.The current research shows that using graphs of neural networks can help researchers learn how to effectively construct representations of academic papers,but most have focused on a supervised learning approach,which requires massive amounts of high quality data.Based on this context,in this study uses four unsupervised graphs of neural networks for learning to construct representations of academic papers using data from the Cora,CiterSeer,and DBLP datasets;then,the representations of the papers are applied to completing two downstream tasks,i.e.,literature classification and paper recommendation.The experimental results show that:(1)deep graph infomax performs best for literature classification,and an adversarially regularized graph autoencoder exhibits better performance in paper recommendation;and(2)citation networks perform better than cocitation networks and literature coupling networks on the Cora dataset. |
| |
| Keywords: | unsupervised learning graph of neural network representation learning academic paper |
| 本文献已被 维普 万方数据 等数据库收录！ |
|
