| 李琳娜,郭晓琪,张运良,王力,张晓丹.基于精细异质信息网络表示学习的文献推荐研究[J].数字图书馆论坛,2025,21(5):11~19 |
| 基于精细异质信息网络表示学习的文献推荐研究 |
| Literature Recommendation Based on Representation Learning of Fine-Grained Heterogeneous Information Network |
| 投稿时间：2024-09-06 |
| DOI：10.3772/j.issn.1673-2286.2025.05.002 |
| 中文关键词: 精细异质信息网络；异质信息网络表示学习；文献推荐；细粒度标签；计算机科学；研究问题；
方法模型 |
| 英文关键词: Fine-Grained Heterogeneous Information Network; Representation Learning of Heterogeneous Information Network; Literature Recommendation; Fine-Grained Label; Computer Science; Research Question; Methodological Model |
| 基金项目:本研究得到中国科学技术信息研究所重点工作项目“面向战略决策的智能情报技术引擎研究及应用”（编号：ZD20250-08）资助。 |
| 作者 | 单位 | | 李琳娜 | 中国科学技术信息研究所；富媒体数字出版内容组织与知识服务重点实验室 | | 郭晓琪 | 中国科学技术信息研究所 | | 张运良 | 中国科学技术信息研究所；富媒体数字出版内容组织与知识服务重点实验室 | | 王力 | 中国科学技术信息研究所；富媒体数字出版内容组织与知识服务重点实验室 | | 张晓丹 | 中国科学技术信息研究所 |
|
| 摘要点击次数: 38 |
| 全文下载次数: 73 |
| 中文摘要: |
| 目前针对异质信息网络的文献推荐主要基于已有的文献间引用、共同作者、同期刊发表等信息构建文献之间的关联关系，没有从文献的研究问题、方法模型等研究人员更关心的文献细粒度内容角度构建文献间关系，不能将这些细粒度的关联关系融入文献推荐过程，从而影响最终的推荐效果。将文献的研究问题、方法模型两类细粒度标签加入学术信息网络，并提出了文献推荐模型PRM-FHIN，通过对异质信息网络表示学习方法HECO进行优化学习网络节点的结构向量，通过对SciBERT模型进行微调学习网络节点的内容向量，基于融合的内容向量和结构向量实现最终的文献推荐。从开放学术图谱中抽取2010—2020年计算机科学领域的185万篇论文作为实验数据，实验结果表明：优化后的HECO算法能更好地对网络节点进行嵌入式表示，在异质信息网络中融入研究问题和方法模型等细粒度标签可以丰富文献之间的语义信息，从而提高最终的文献推荐效果。 |
| 英文摘要: |
| Current research on literature recommendation based on heterogeneous information network are primarily based on existing citations, coauthorships,and publications in the same journals to build the relationship between literature. These methods do not build relationships from the perspectives of the research questions and methodological models of the literature, which are the finer-grained contents that researchers are more concerned about. This limitation prevents the integration of these fine-grained relationships into the literature recommendation process, thereby affecting the final recommendation effectiveness. This paper incorporates two types of fine-grained labels, namely research questions and methodological models of literature, into an academic information network and proposes a literature recommendation model called PRM-FHIN. It optimizes the HECO approach for heterogeneous information network representation learning to acquire structural vectors of network nodes and fine-tunes the SciBERT model to learn content vectors of network nodes. The final literature recommendation is achieved based on the integration of these content vectors and structural vectors. Using 1.85 million papers in the field of computer science from 2010 to 2020 extracted from the Open Academic Graph as experimental data, the results show that the optimized HECO algorithm can better embed the network nodes. Incorporating fine-grained labels such as research questions and methodological models into the heterogeneous information network enriches the semantic information between the literature, thereby improving the final literature recommendation effectiveness. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
