【作者】 王鑫；

【导师】 赵姝； 陈洁；

【作者基本信息】 安徽大学 ， 计算机技术（专业学位）， 2020， 硕士

【摘要】 面对综合化、多元化、交叉化的复杂科学研究,学者之间的学术合作使得在不同领域的学者可以学习新知识、拓宽新视野、实现科研资源的高效利用,从而加快科研信息的流动,提升科研成果的质量,进而对解决技术难题和理论创新具有深刻意义。合作者推荐是为学者推荐合适研究领域的学者进行相关学术合作。然而,许多研究仅通过文本挖掘技术得到学者的文本特征表示或是通过合作关系网络得到学者的结构特征表示以及通过文本特征表示和结构特征表示简单融合得到学者的特征表示,而忽略了文本信息和网络结构相互促进对学者的特征表示产生的影响。网络表示学习旨在将网络中的节点表示成低维向量,而学术合作网络中包含许多文本信息和网络结构信息,这些信息可以精细化网络结构,有利于学者的特征表示。无论是基于关键词检索还是基于主题模型的专家档案建立,不同的文本表示之间会存在语义鸿沟问题,而学者文本特征和网络结构特征简单的融合相加不能准确捕获学者之间的潜在关系。本文将学者文本信息和网络结构信息相融合,利用网络表示学习技术挖掘学者的特征表示,帮助学者推荐合适的合作者。本文的研究工作主要包括:1、针对合作关系网络中蕴含的潜在合作关系,提出一种基于属性网络表示学习和主题感知的学术合作者推荐(TACR-ANRL)。通过构建由论文、学者和主题组成的异构信息网络捕获合作关系网络中学者尽可能的潜在合作关系,利用属性网络表示学习将学者的文本信息和网络结构映射到同一向量空间来捕获学者之间的潜在合作关系,从而得到精确的学者特征表示,最后基于学者的特征表示计算学者之间的相似度为每个学者推荐top-k个相似的合作者。在数据集上进行实验,结果表明本研究能够有效的利用主题捕获合作关系网络中由学者文本信息和网络结构融合产生的潜在合作关系并且相比较于其他合作者推荐方法能取得良好的表现。2、针对属性组成的语义关系对合作关系的增强特征,提出一种基于文本增强型网络表示的学术合作者推荐(CNEacR)。该方法主要利用学者文本信息得到学者的加权文本表示,接着构建一个既包含原生合作关系又包含由文本增强的语义关系形成的文本增强型的学术合作者网络,通过网络表示学习得到学者的特征表示,最后基于学者的特征表示计算学者之间的相似度为每个学者推荐top-k个相似的合作者。在数据集上进行实验,结果表明本研究能够有效的捕获合作关系网络中学者文本信息和网络结构融合产生的潜在合作关系并且相比较于其它合作者推荐方法能取得良好的表现。更多还原

【Abstract】 In the face of comprehensive,diversified and intertwined complex scientific research,academic collaborations among scholars enable scholars in different fields to learn new knowledge,broaden new horizons,and realize the efficient use of scientific research resources,accelerating the flow of scientific research information,improving the quality of scientific research,which has profound significance for solving technical problems and theoretical innovations.The collaborator recommendation is for scholars to recommend scholars in suitable research fields for relevant academic collaborations.However,many studies only obtain the scholar’s text feature representation through text mining technology or through the cooperative relationship network to obtain the scholar’s structural feature representation and through simple integration of text feature representation and structural feature representation to obtain the scholar’s feature representation,however,the influence of text information and network structure on the feature representation of scholars is ignored.Network representation learning aims to represent nodes in the network as low-dimensional vectors,while the academic cooperative network contains a lot of text information and network structure information.This heterogeneous information can refine the network structure,which is beneficial to the feature representation of scholars.Whether keyword retrieval or expert archives based on the topic model,there will be a semantic gap between different text representations,and the simple fusion of scholars’ text representation and network structure representation can not accurately capture the potential relationship between scholars.In this dissertation,the text information of scholars is fused with the network structure,and the feature representation of scholars is mined by using the network representation learning technology to help scholars recommend proper collaborators.The main work of this dissertation is as follows:There are many potential partnerships in the partnership network,this dissertation proposes a method that Topic-aware Academic Collaborator Recommendation based on Attributed Network Representation Learning,which is called TACR-ANRL for short.By building a heterogeneous information network composed of papers,scholars,and topics to capture the potential collaboration relationships of scholars in the collaboration relationship network as much as possible,and using attribute network representation learning to map scholars’ text information and network structure to the same vector space to capture potential cooperative relationships between scholars,we obtain an accurate feature representation of scholars.Finally,based on the scholar’s feature representation,the similarity between the scholars is calculated to recommend top-k similar collaborators.Experiments on the data set show that this method can effectively use the topic to capture the potential collaborator relationships generated by the fusion of scholar text information and network structure in the collaborator relationship network and can achieve good performance compared with other collaborators’ recommendation methods.The nodes in the collaborator relationship network have rich attributes,and the semantic relationship composed of the attributes can be used to enhance the network structure of the collaborator relationship.This dissertation proposes a method that Content-enhanced Network Embedding for Academic Collaborator Recommendation,which is called CNEacR for short.This method mainly uses the scholar’s text information to obtain the scholar’s weighted text representation and then constructs a content-enhanced collaborator network that contains both the original collaborator relationship and the content-enhanced semantic relationships,and the scholar’s feature representation is obtained through the network representation learning.Finally,calculating the similarity based on the feature representation of scholars between scholars recommends top-k similar collaborators for each scholar.Experiments on the data set show that this method can effectively capture the potential collaborator relationships generated by the fusion of scholar text information and network structure in the collaborator relationship network and can achieve good performance compared with other collaborator recommendation methods.更多还原