基于合作模式的学者影响力预测研究

发布时间：2019-01-20 08:38

【摘要】：随着学术社会的不断发展,近年来学术论文的产量不断增长,学术信息不断丰富,学术大数据逐渐成为一个新兴的研究领域。学术影响力评估是学术社会中必不可少的一部分。在学术社会中,我们可以通过选择跟踪有影响力的学者,获得某一领域的最新研究进展;在学校职称评估中,我们可以通过学者的学术影响力来决定职称的评定;在基金申请中,往往具有高影响力的学者可以更有可能获得资助。然而,随着大数据时代的来临,学术信息过载问题逐渐成为学术合作的壁垒。即,学者们很难权衡一个学者的影响力。因此,如何能够有效的获取和预测一个学者的影响力是一个亟待解决的问题。为了解决学术大数所带来的问题和挑战,研究者们从不同角度对各种不同学术关系进行研究和分析。在本文中,我们主要研究学术大数据中,学者的影响力预测问题,即给定一个学者,预测其在若干年后的学术影响力。这有助于了解学者的研究能力和公正的评价一个学者的学术能力,从而为基金分配,职称评定等等实际问题提供帮助,解决学术信息过载问题。同时,深度学习,是近些年机器学习研究中的一个领军方向,深度学习的核心思想是,通过分析和模拟人脑在处理数据时的方式方法,通过模仿人类神经网络的思考方式来处理数据。在数据量庞大的信息化时代,深度学习在预测领域也可发挥其巨大的作用。同时,区别于以往的基于论文引用的学术影响力预测方法,我们提出基于合作模式的学术影响力预测方法。本文不同于传统的学术影响力预测技术,提出一种基于深度学习的学术影响力预测方法。通过对深度学习方法的学习和深入研究,将其引入到传统的预测算法中。在基于学者合作模式的学者影响力预测技术中,我们采用了自动编码器(Stacked Autoencoder),提出了基于合作模式的学术影响力预测模型SIP(Scholar Impact Prediction),包括基于学术合作网络的特征提取,和基于编码器的影响力预测两个部分。首先,基于学术合作网络,我们提取学者的个人属性和网络属性等学者合作模式作为深度学习训练的输入特征。然后根据得到的属性特征集,利用自动编码器算法学习模型参数来进行学术影响力预测。本文在DBLP学术数据对所提出的SIP算法进行评估,结果表明,与传统的机器学习算法对比,我们提出的基于合作模式的学术影响力预测模型在平均绝对误差,均方根误差和皮尔逊相关系数等指标上表现较好,表明了深度学习在学术影响力预测的性能。同时,我们分析了各个特征属性对预测结果的影响,可以帮我们更加深刻的理解不同学者合作模式对预测结果的影响。另一方面,本文探索的学术社会中的影响力问题,在现实生活中,在不同的在线社交网络中,也存在同样的问题,由于网络属性的不同,预测技术各有不同,同时彼此相通。因此,我们所提出的预测方法,具有一定的普适性,也具备应用到其他预测或者推荐系统之中。例如,电子商务中的物品推荐,在线社交媒体中的还有推荐,以及基于兴趣的物品推荐等等。因此,我们所提出的预测策略具有一定的普适性,对其他预测技术有一定的借鉴意义。
[Abstract]:With the development of the academic society, in recent years, the output of the academic papers has been growing, the academic information is constantly enriched, and the large-scale academic data has gradually become a new field of research. The evaluation of academic influence is an essential part of the academic society. In the academic society, we can obtain up-to-date research progress in a certain field through the selection of influential scholars; in the assessment of the title of the school, we can determine the assessment of the professional title through the academic influence of the scholars; in the application of the fund, Scholars who tend to have a high impact can be more likely to be funded. However, with the advent of large data age, the problem of academic information overload has become a barrier to academic cooperation. That is, it is hard for scholars to weigh the influence of a scholar. Therefore, how to effectively acquire and predict the influence of a scholar is an urgent problem to be solved. In order to solve the problems and challenges brought by the large number of academic studies, the researchers have studied and analyzed various academic relations from different angles. In this paper, we mainly study the influence prediction of the scholars in the academic big data, that is, given a scholar, the academic influence of the scholar after several years is predicted. This will help to understand the research ability of the scholars and the impartial evaluation of the academic ability of a scholar, so as to provide help to the practical problems such as fund distribution, job evaluation and so on, and solve the problem of the overload of the academic information. At the same time, the depth study is a leading direction in the study of machine learning in recent years. The core idea of depth study is to process the data by analyzing and simulating the way the human brain processes the data. In the era of large data volume, depth study can also play a great role in the field of prediction. At the same time, it is different from the previous academic influence prediction method based on the paper reference, and we put forward the method of predicting the academic influence based on the cooperative model. This paper is different from the traditional academic influence prediction technology, and puts forward a method of academic influence prediction based on depth study. The learning and in-depth study of the depth learning method is introduced into the traditional prediction algorithm. In the field of the scholars' influence prediction based on the cooperative model of the scholars, we have adopted the Stamped Autoencer, and put forward the model of academic influence prediction (SIP) based on the cooperation model, including the feature extraction based on the academic cooperation network. and predicting two parts based on the influence of the encoder. First, on the basis of the academic cooperation network, we extract the personal attributes and the network attributes of the scholars as the input feature of the depth study training. and then using the automatic encoder algorithm to study the model parameters to predict the academic influence according to the obtained attribute set. In this paper, the proposed SIP algorithm is evaluated by the DBLP academic data. The results show that, compared with the traditional machine learning algorithm, the model of the academic influence based on the cooperative model is better than the average absolute error, the root mean square error and the Pearson correlation coefficient. The performance of depth learning in the prediction of academic influence is shown. At the same time, we analyze the effect of each feature attribute on the prediction result, and can help us to understand the effect of different scholars' cooperation model on the prediction result. On the other hand, the influence of this paper in the academic society, in the real life, in different online social networks, there are also the same problems, because of the different network attributes, the prediction techniques are different, and at the same time they are in communication with each other. Therefore, the prediction method proposed by us has a certain universality and is also applied to other prediction or recommendation systems. e. g., an item recommendation in e-commerce, a further recommendation in an online social media, and an interest-based item recommendation, and the like. Therefore, the prediction strategy proposed by us has a certain universality, which can be used for reference for other prediction techniques.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181

【相似文献】