【作者】 刘鑫；

【导师】 查宏远；

【作者基本信息】 华东师范大学 ， 软件工程（专业学位）， 2018， 硕士

【摘要】 科技文献包含了期刊论文、会议论文、学位论文、发明专利以及科技报告等。随着科学技术的快速发展,科技文献数量急剧增长,科学而准确地评估文献当前以及潜在的影响力具有重要的社会价值,例如在科学基金资助、专利诉讼、人员聘用等方面。而文献的影响力和文献相互间的引用行为有着密切的联系,这使得对文献引用行为的研究成为有价值的学术课题。本文在充分考量文献引用行为数据的特性的基础上,提出了一个合理的多维时间点过程模型,将其应用于科技文献引用行为数据的建模和分析;基于极大似然和EM估计方法,本文设计和实现了高效率的预训练算法,用以快速估计点过程模型参数;本文还将判别式学习方法、生成对抗网络方法与时间点过程结合,设计了完整的全训练算法,进一步挖掘点过程模型的预测潜力:本文在真实的学术论文和科技专利的引用数据上进行了实验分析和探讨,其中学术论文数据来源于微软学术图(MAG)[1],而科技专利数据来源于NBER[2];此外,本文还探讨了学术搜索预测引擎的工程实现,通过实际的工程项目证实本文所设计的模型的合理性。本文所提出的模型可实现预估未来任意一段时间内的引用,实现了对文献影响力的预测性分析,而不再局限于完全基于过往引用状况的分析。本研究设计了衡量文献中长期影响力的辅助工具,为文献潜力的评估提供了科学的参考,推动机器学习方法在这一领域的应用实践,是人工智能技术发挥实际社会价值的实例。同时,本文所设计和实现的点过程模型的应用领域并不局限于科技文献的引用行为建模,它在金融投资行为建模、企业并购行为建模、职业发展建模等其他类似问题中都有应用潜力。更多还原

【Abstract】 Categories of scientific literature include journal,conference paper,patent,techni-cal report,dissertation,etc.The literature is turning into an unbounded collection as the rapid development of science and technology,it is of great importance to evaluate the cur-rent and potential impact as accurately as possible in fields such as personnel recruitment,funding,patent infringement lawsuit.It is widely believed that the value of literature is often closely related to peer-citations,which makes research on citation behavior an im-portant topic.On the basis of specific insight on literature citation data,this work attempts to model the peer-citation behavior,proposing a multi-dimensional temporal point process model;Based on maximum likelihood evaluation(MLE)and EM method,we have de-signed and implemented a highly efficient evaluation algorithm to train this model rapidly;furthermore,we have applied discriminative and adversarial learning method to devise a full training procedure,in order to further enhance prediction ability of point process;we conduct experiments on real citation data of academic paper data(originating from Microsoft Academic Graph[1])and patent data(originating from NBER[2]);besides,we have also discussed the implementation of an academic search and prediction engine.The proposed model is capable of evaluating future citations in an arbitrary period,thus en-abling the predictive analysis for literature’s impact.This work designs an effective tool for quantifying long-term scientific and technical impact,offering suggestions for liter-ature’ s potential value,promoting the application of machine learning method in this field,and is an instance that AI technology plays an importance role in real life problems.Meanwhile,proposed model can also be applied in fields such as high frequency trading,merger and acquisition,career path modeling,etc.更多还原