融合异质图表示学习与注意力机制的可解释论文推荐  

Explainable Paper Recommendations Based on Heterogeneous Graph Representation Learning and the Attention Mechanism

作　　者：马霄 邓秋淼 张红玉 文轩 曾江峰 Ma Xiao;Deng Qiumiao;Zhang Hongyu;Wen Xuan;Zeng Jiangfeng(School of Information Engineering,Zhongnan University of Economics and Law,Wuhan 430073;School of Information Management,Central China Normal University,Wuhan 430079)

机构地区：[1]中南财经政法大学信息工程学院,武汉430073 [2]华中师范大学信息管理学院,武汉430079

出　　处：《情报学报》2024年第7期802-817,共16页Journal of the China Society for Scientific and Technical Information

基　　金：国家自然科学基金青年科学基金项目“情感感知的可解释虚假新闻检测研究”(62102159);“基于多源异质信息融合的推荐算法研究”(61802440);教育部人文社会科学研究青年基金项目“情境大数据驱动的社交媒体虚假信息识别模型与治理策略研究”(21YJC870002);湖北省自然科学基金项目“基于知识图谱增强的个性化联邦推荐方法研究”(2024AFB957),“基于多层语义融合的多模态社交媒体虚假信息检测研究”(2023AFB1018)。

摘　　要：学术论文推荐旨在为研究人员从海量学术资源中快速筛选出感兴趣的论文。现有论文推荐方法主要基于论文标题等文本内容和引用关系等进行推荐,使得蕴含丰富语义的多源学术信息的表示学习不够充分,制约了推荐准确度的进一步提升。同时,当前方法往往关注论文推荐的准确性,而忽略了可解释性,降低了论文推荐系统的可信度和用户满意度。为解决上述问题,本文提出了一种融合异质图表示学习与注意力机制的可解释论文推荐方法,该方法能够有效利用异质学术图中的语义信息,为推荐结果提供文本解释说明。具体来说,首先,提出了一种基于异质图表示学习与注意力机制的论文推荐模型,融合多源学术信息来构建语义丰富的异质学术图,并利用注意力机制学习不同节点和元路径的重要性,以获得更准确的节点表示。其次,提出了一种基于特征的文本解释生成模型,该模型将可解释文本生成方法引入论文推荐场景,能够在为作者提供推荐列表的同时生成文本解释,以告知其推荐缘由,从而提高论文推荐的可解释性。最后,构建了一个包含论文元数据、特征词、引用上下文的学术数据集,基于该数据集的对比实验结果表明,本文提出的基于异质图表示学习与注意力机制的论文推荐模型推荐准确度更高,解释生成模型能够为论文推荐结果提供质量较高的可解释文本说明。Paper-recommender systems aim to recommend the most relevant academic papers for researchers from the vast academic resources.Existing paper recommendation methods are mainly based on the textual information of papers or citation relationships,failing to make full use of the rich semantic information in heterogeneous academic graphs,thus leading to less accurate recommendation results.In addition,existing methods focus more on recommendation accuracy and neglect interpretability, which decreases the reliability and user satisfaction of the paper recommender systems. To solve these issues, this study proposes an Explainable Paper Recommendation (EPRec) method based on heterogeneous graph representation learning and the attention mechanism. First, an attention mechanism-enhanced heterogeneous graph representation learning-based paper recommendation module is proposed to incorporate multi-source side information from heterogeneous academic graphs. Then, the interpretable text generation method is introduced into the paper recom‐mendation scenario. A feature-based text generation module is proposed to generate textual interpretation to explicitly pro‐vide recommendation reasons to researchers. Finally, we construct an academic dataset that contains multi-source informa‐tion, including paper metadata, feature words, and citation contexts. The experimental results show that EPRec performs better than the comparative methods in terms of precision and recall. Moreover, EPRec provides high-quality, interpretable text explanations for the paper recommendation results.

关 键 词：论文推荐 异质图表示学习 注意力机制 可解释文本生成 元路径 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]