异构学术网络多关系普适的链接预测模型研究

发布时间：2018-05-22 17:15

本文选题：异构学术网络 + 链接预测　；参考：《北京邮电大学》2017年硕士论文

【摘要】：随着科学研究进程日益加快,学术资源的数量呈现爆炸式增长。面对庞大的学术资源,科研人员希望能够快速准确地找到与其研究相关的论文、作者,或者可以投稿的期刊等学术实体。为应对上述挑战,越来越多的研究者从不同的角度研究学术网络。其中通过构建同构或异构的学术网络,利用链接预测方法对缺失或未来链接进行预测,对于实体推荐、关系预测及网络分析等目标具有重要意义。当前异构学术网络的链接预测方法仅能预测单一或局限的学术实体间关系,构造的链接预测模型普适性较差,链接预测的准确性有待提高。针对上述问题,本文提出了一种异构学术网络多关系普适的链接预测方法。该方法构建了一个包含当前常见学术实体类型及实体间关系类型的异构学术网络,提出了一种能够挖掘深层网络特征的自动化元路径提取方法,并提出了两种改进的融合时间和实体间内容相关性的元路径相似性度量方法,以对任意类型的元路径进行度量,最后选择多关系普适的逻辑回归模型进行特征参数学习来构建最终的链接预测模型。本文使用微软学术数据作为实验数据集构建异构学术网络,基于本文提出的普适链接模型对网络中具有实际含义的多种关系类型进行了自动化的链接预测。实验结果表明,本文采用统一的链接预测模型可以进行多种链接关系的自动化预测,链接预测准确率得到明显提高,同时通过时间扩展实验验证了本文方案随时间推移仍然有较好的适用性。
[Abstract]:With the rapid development of scientific research, the number of academic resources is increasing explosively. Faced with huge academic resources, researchers hope to find papers, authors, and journals that can contribute to their research quickly and accurately. To meet these challenges, more and more researchers are studying academic networks from different angles. Among them, it is important for entity recommendation, relationship prediction and network analysis to predict missing or future links by constructing isomorphic or heterogeneous academic networks. The current link prediction methods of heterogeneous academic networks can only predict the relationship between single or limited academic entities, and the constructed link prediction model is of poor universality, and the accuracy of link prediction needs to be improved. In order to solve the above problems, this paper presents a link prediction method for heterogeneous academic networks. This method constructs a heterogeneous academic network which includes the common types of academic entities and the types of relationships between entities, and proposes an automated meta-path extraction method which can mine the features of deep network. Two improved meta-path similarity measurement methods are proposed to measure the meta path similarity of any type of meta path, which combines time and content correlation between entities. Finally, the multirelational universal logical regression model is chosen to study the feature parameters to construct the final link prediction model. In this paper, we use Microsoft academic data as experimental data set to construct heterogeneous academic network. Based on the universal link model proposed in this paper, we make an automated link prediction for many kinds of relationships with actual meaning in the network. The experimental results show that the unified link prediction model can be used to predict many kinds of link relationships automatically, and the accuracy of link prediction is improved obviously. At the same time, the experimental results of time extension show that the proposed scheme has good applicability over time.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：O157.5

【相似文献】