科研社交网站中的学者推荐研究

发布时间：2018-11-08 16:26

【摘要】：Web2.0时代,社交网络用户可以自由的发布信息、交流思想,吸引了人们在该类平台上建立社区、交流知识,由于一般社交网络缺乏专业学术氛围,2007年起出现了专门面向学术工作者的科研社交网站,如国外的ResearchGate、Academia.edu,国内的百度学术、科研之友等。他们在网站中浏览彼此主页、寻找感兴趣的文献与学者、参与学术话题讨论、相互提问解答,这使得全球各领域科研人员能够方便地进行即时学术探讨、寻求潜在合作机会。发现相似研究学者与潜在合作者是科研工作者使用网站的重要理由之一。但是,科研社交网络存在与大众社交网络相同的信息过载、信息不对称的问题,基于学者的学术知识与科研合作网络构建个性化推荐模型是有效的解决手段。进一步地,目前信息处理与检索系统的一个新趋势是对情境化数据的获取,将其考虑进信息处理中,有助于提高推荐精确度,缓解信息过载,更好的适应与用户已有历史记录相独立的特殊需求。为此,本文分析了科研社交网站中学者的社交动机,得出推荐场景差异,认为学者主要对同一研究领域、具有相似研究偏好的学者感兴趣,并与他们建立长期的社交关系,除此外,很多学者具有情境化特征,希望寻找具有特定要求限制下的合作者,如已有研究主题的项目或者论文。因此,本文提出了两个学者推荐模型,即基于相似研究兴趣的学者推荐模型,和基于特定情境的合作者推荐模型。针对两种推荐情境,本文分别设计了合理对应的解决策略。在基于相似研究兴趣的学者推荐模型中,本文构造了两个子模型:学者档案模型与学术行为网络模型。在学者档案模型中,采用语言模型,依据学者的专业、研究领域、研究成果等信息表征学者知识,使用基于贝叶斯分解的生成概率计算学者知识的相似度;在学术行为网络模型中,通过挖掘学者学术行为网络中的关系,采用Adamic-Adar方法和最短路径方法分别测量合作者网络中的学者节点相似度和路径距离,从全局学术领域和局部研究领域两个角度采用Jaccard系数表示研究学者所在单位间的合作网络关系度;最后,应用Comb策略整合以上测量,预测相似度较高的学者为推荐学者。在基于特定情境下的合作者推荐模型中,本文设计了两个标准评定潜在合作者的质量:学者学术质量评价与学术社会网络质量评价。在学者学术质量评价中,同时引入情境预过滤和情境后过滤到推荐方法中,使用学者的学术成果质量(成果数量、发表刊物级别、被引用量)、职称、G指数来为学者的学术能力评分,对情境信息进行预处理、提取特征词,首先采用情境预过滤策略选出含有情境内容特征的学者构成初步候选合作者集,然后采用调整的潜狄利克雷分配方法对情景主题分配关键词,运用Kullback-Leibler差异计算初步候选合作者集中的学者与目标学者间的知识匹配,并将MNZ标准化后的学者学术能力评分作为匹配计算中的权重值;在学术社会网络质量评价中,构建了多元关系网络,包括四种关系类型:论文合作、项目合作、专利合作、出席相同会议,先计算学者间四种关系的数量,再引入关系年限修正得到合作质量评分;最后对两项评分进行整合得到合作意向评分。两个推荐模型的具体构建方法见于论文第四章。同时,为了模型应用的清晰与完整性,本文构建了科研社交网络的全局系统架构,并采用Python2.7+Selenium+Scrapy搭建爬虫,得到2000名学者作为模型模拟数据集,收集到共8万多项学术成果,本文在第五章中详细给出了模型的应用过程与每一步骤的计算值,实验结果表明,模型具有应用可行性和较好的推荐效果。
[Abstract]:In the Web 2.0 era, social network users can freely distribute information and exchange ideas, and attract people to build community and exchange knowledge on such platforms. As a result of the lack of professional academic atmosphere in the general social network, social networking sites for academic workers have emerged in 2007. such as the foreign research group, the Academia.edu, the domestic Baidu academic, the scientific research friends and so on. They visit each other's home page in the website, find the literature and scholars of interest, participate in the discussion of the academic topic, and ask each other to answer the questions, which makes the researchers in all fields of the world can carry on the real-time academic discussion and seek the potential cooperation opportunity. It is found that similar research scholars and potential collaborators are one of the important reasons for scientific research workers to use the website. However, the social network of scientific research has the same information overload and information asymmetry with the public social network, and it is an effective solution to construct the personalized recommendation model based on the academic knowledge of the scholars and the scientific research cooperation network. Further, a new trend of the present information processing and retrieval system is to obtain the contextual data, to take into account the information processing, to improve the recommended accuracy, to alleviate the overload of the information, and to better adapt to the special needs independent of the user's existing history. To this end, this paper analyzes the social motivation of the scholars in the social network of scientific research, and draws the difference of the recommended scene, and thinks that the scholars are interested in the same research field and have similar research and preference, and establish a long-term social relationship with them, and, in addition, Many of the scholars have a contextual feature, looking for collaborators with specific requirements, such as projects or papers with research topics. In this paper, two academic recommendation models, namely, a recommendation model based on similar research interests, and a co-author recommendation model based on a specific context, are proposed in this paper. In this paper, a reasonable and corresponding solution is designed for two recommended situations. In the author's recommendation model based on similar research interest, the paper constructs two sub-models: the academic file model and the academic behavior network model. in that scholar's file model, the language model is adopted, and the knowledge of the scholar is characterized by the information of the professional, the research field, the research result and the like of the scholar, and the similarity of the knowledge of the scholar is calculated by using the generation probability based on the Bayesian decomposition; in the network model of the academic behavior, by digging the relationship between the scholar's academic behavior network, the Adamic-Adar method and the shortest path method are adopted to measure the similarity and the path distance of the scholar nodes in the partner network, From the global academic field and the local research field, the Jaccard coefficient is used to show the relationship degree of the cooperative network between the research scholars and the research scholars. Finally, the author uses the Comb strategy to integrate the above measurement, and the scholars with higher similarity are predicted to be the recommended scholars. In the model of co-author's recommendation based on the specific situation, this paper designs two criteria to evaluate the quality of potential collaborators: the evaluation of the academic quality and the evaluation of the network quality of the academic society. in that evaluation of the academic quality of the scholar, the situation pre-filtration and the situation are introduced into the recommendation method, the academic achievement quality (the number of achievement, the publication level, the quoted amount), the professional title and the G index of the scholar are used to score the academic ability of the scholar, pre-processing the situation information, extracting the characteristic words, first adopting a context pre-filtering strategy to select a first candidate partner set containing the context content feature, and then using the adjusted latent Dirichlet distribution method to assign a keyword to the scene theme, using the Kullback-Leibler difference to calculate the knowledge matching between the scholars and the target scholars in the preliminary candidate collaborator, and the academic ability score of the scholars after the MNZ standardization is used as the weight value in the matching calculation; in the network quality evaluation of the academic society, a multi-element relationship network is constructed, It includes four types of relationship: paper cooperation, project cooperation, patent cooperation, attendance at the same meeting, first calculating the number of four relationships among the scholars, and then introducing the relationship life correction to obtain the cooperation quality score; and finally, combining the two scores to obtain the cooperation intention score. The specific construction methods of two recommended models are found in the fourth chapter of the thesis. At the same time, for the sake of clarity and completeness of the model application, this paper constructs the global system architecture of the scientific research social network, and sets up the crawler by Python2.7 + Selenium + Srapy, and obtains 2000 scholars as the model model data set, and collects over 80,000 academic achievements. In the fifth chapter, the application process of the model and the calculation value of each step are given in detail. The experimental results show that the model has the application feasibility and the better recommended effect.
【学位授予单位】：华中师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP393.092;G354

【参考文献】