【作者】 罗鹏；

【导师】 吴冲；

【作者基本信息】 哈尔滨工业大学 ， 管理科学与工程， 2019， 博士

【摘要】 复杂网络模型是对现实世界复杂系统的结构化和简单化,以复杂系统中的实体抽象为节点,实体关系抽象为链接,通过分析网络结构或者网络影响来解释现实世界的现象。本文主要研究网络中的“同伴”效应,考虑目标节点周围有紧密关系(直接链接或者间接链接)的节点影响。目前有关复杂网络“同伴”效应的研究中,学者们主要探究了以人为参与者的社交关系网络中“同伴”的影响,在以客观实体为参与者的实体网络(例如产品关系网络和企业关系网络)中,仅仅分析了网络结构的影响,忽略了“同伴”效应的作用。另外,现有研究中“同伴”效应一般都是关注直接链接“同伴”的影响,忽略了网络中更多间接链接“同伴”的作用。针对以上研究空白,本文进行了如下的复杂网络中的“同伴”效应的研究。首先,本文选择在线医疗社区作为第一个研究对象,以在线医疗社区中同一医院同一科室内部医生的工作关系构建无向无权重的医生网络,选择医生在医疗社区开放的可访问渠道数目作为因变量(代表了医生参与在线医疗社区的程度),“同伴”效应指同一医院同一科室中目标医生周围其他同事开放的可访问渠道数目的均值。通过收集实际数据,本文实证了“同伴”效应对目标医生的行为具有积极促进作用。另外,针对同一科室内部,医生之间不仅存在合作促进关系,还存在竞争关系,促进关系可以由“同伴”效应的直接正向作用体现,为了研究竞争关系,本文探究“同伴”效应对于医生评价和因变量关系的调节作用,结果显示“同伴”效应的调节作用是负向显著的。其次,为了进一步研究客观实体网络中的“同伴”效应,本文收集了社交媒体数据,通过文本分析和情感分析方法识别企业之间的比较关系,构建了无向有权重的企业共现网络和有向有权重的企业比较关系网络。另外,利用百度搜索数据作为企业影响力数据。根据构建的企业网络模型,本文定义了“同伴”数量和加权“同伴”影响两个指标,并利用向量自回归模型(VAR)分析了“同伴”效应指标与企业股票收益和风险的动态关系。实证结果表明,企业关系网络中,“同伴”效应指标能够显著影响企业的收益和风险,同时证明了比较关系网络“同伴”效应指标对于企业股票表现(收益和风险)的解释能力要显著大于企业共现网络“同伴”效应指标。另外,为了验证网络中间接链接的“同伴”效应,本文收集了产品评论数据,从产品评论数据中,识别产品之间的比较关系,同时结合文本情感分析方法,构建了上层有向有权重的产品比较关系网络。另外,本文还考虑了用户在产品评论的交流关系,通过用户评论和回复的结构,构建了下层有向有权重的用户评论网络。在构建的多层次产品和评论网络模型的基础上,定义了网络中传递的“同伴”效应指标。结合产品销售排名数据,我们利用回归模型和分类模型,研究传递的“同伴”效应指标对于产品销售排名的作用。实证结果表明:传递的“同伴”效应指标对于产品销售排名具有显著作用,同时对于因变量的解释能力和预测能力都是最强的。最后,本文将“同伴”效应这一想法应用到网络链路预测中。通过构建网络节点的效用模型,本文考虑不同距离的“同伴”效应建立了四个不同的模型。结合现有网络状态,利用二元逻辑回归模型,对模型的参数进行估计,最后得到四个链路预测模型。结合社交网络数据,利用我们构建的链路预测模型与现有模型进行比较分析,验证了模型的有效性。另外还把链路预测模型应用到分析社交网络结构的演化中,并进行了朋友推荐的应用。本文通过收集实际数据,构建了新兴的在线医疗社区中的人际关系网络模型,还包括企业、产品等客观实体为组成部分的客观实体网络模型。在不同的网络模型中,定义了仅考虑直接链接的“同伴”效应和同时考虑直接链接和间接链接的“同伴”效应,实证了网络中的“同伴”效应对人的行为,企业股票表现(收益和风险)和产品销售排名都具有显著的影响。另外,本文还考虑网络中直接链接和间接链接的“同伴”效应建立了网络链路预测模型,验证了模型的有效性,并应用于分析网络结构演化和朋友推荐领域。本研究不仅对“同伴”效应和复杂网络相关文献具有一定的理论贡献,同时也具有一定的实际应用价值。更多还原

【Abstract】 The complex network model is the structuring and simolication of the complex systems that the entities are represented by nodes and relationships are represented by links.Scholars can analyze the network structure and network influence to illustrate the real world.This thesis focus on the peer effect nearby the target node.In existing peer influence literature,scholars mainly focus on friendship network and ignore the objective entity network including firm network and product network.The previous studies only examine the influence of network structure in the firm network or product network.Furthermore,as to the peer effect in social network,it only investigates the individuals which are linked to the target person and ignores the individuals which are not directly linked to the target one.Thus,this thesis focuses on the peer effect in complex networks,especially in obejective entity network(firm network and product network),and also considers the peer effect in application research such as link prediction method.Firstly,the new online platform is chosen,i.e.,online health community,to examine the peer effect.The number of channels to access the doctor is adopted as our dependent varaibels representing the participation degree of doctors in onli ne community.The peer effect is defined as the average number of accesss ways of target doctor’s colleagues in the same department of the same hospital.With data from online health community,the results prove the significantly positive relationship between the peer effect and dependent variabel.Furthermore,in the same department,there is not only cooperation relationship but also competitive relationship existed among doctors.The cooperative relations is reflected by the postive effect of peer effect.In order to study the competitive relations,I try to examine the moderation effect of peer influence in the realtionship between doctor rating and dependent variable.The result shows the significant and negative moderation effect of peer influence.Then,to study the peer effect in objective entity network,I collect the social media data including the news data and baidu index data.Based on the text mining method and sentiment analysis method,I construct the firm co-mention network which is undirected and weighted and firm supporting or opposing network which directed and weighted.The thesis uses the baidu index as the firm influence measurements.With the firm network model,I define two variabels including peer number and weighted peer effect.The vector auto-regression model(VAR)is adopted to study the dynamic effect of peer influence variables and the firm stock performance including abnormal return and risk.The empirical results show the significant power of peer effect measurements in illustrating the firm performance.Compared with the variables in firm co-mention network,there is a stronger effect of peer effect metrics in supporting and opposing networks in predicting the firm performance.Next,I further to consider the effect of indirect links,i.e.,the influence of peers which are not directly linked to the target node but can access the target node.I collect the product review data and construct the multi-level network model including the product comparison network and users inter-communication network.In the product comparison network,I adopt the text comparative analysis to identy the product comparative relationship and use the sentiment analysis method to obtain the relationship weight.Moreover,I also consider the users communication structure in product review and construct the user inter-communication network.Based on our multi-level product network model,I construct the transitive peer effect which includes the influence of nodes linked to the target and nodes not linked to the target node but can access the target node.With the product sales ranking data,I use the regression and classification methods to exmaine the impact of transitive peer effect in affecting the product sales ranking.The empirical results show the significant influence of transitive peer effect and prove the strongest ability of transitive peer effect in illustrating the product sales.Lastly,I also apply the peer effect in construct link prediction methods in complex network.With the construction of node utility,I cosider the effects of different peers and construct four different link prediction models.With the existing network state,I use the logistic regression model to estimate the coefficients in the models.Then,with the real social network data,I compare the performance of our proposed link prediction methods and the existing link prediction methods.The results show the superiority of our proposed methods.I also apply our proposed link prediction methods in analyzing the evolution of social networks and friends recommendation.With collecting real data,this thesis examines the peer effect not only in the friendship network such as online health community but also in objective entity network including the product network and firm network.In friendship network and firm network,I construct the peer effect measurements including the influence of peers which are linked to the target node.While in the product network and link prediction methods,except the nodes directly linked to the target one,I also consider influence of nodes which have access to the target node.In this thesis,I prove the significant influence of peer effect in affecting the doctor behavior,firm performance in stock market and product sales ranking.More,I also construct the link prediction methods considering the peer effect and prove the validity of our methods.Our thesis not only contributes to the existing social influence and complex literature,but also provides some insights in the practical implications.更多还原