【作者】 胡艳丽；

【导师】 张维明；

【作者基本信息】 国防科学技术大学 ， 管理科学与工程， 2011， 博士

【摘要】 互联网和Web2.0技术在全球范围内的迅猛发展引发了一场影响深远的媒体革命！网络以其表达的自由性、匿名性、交互性和跨时空性等特性为社会成员提供了空前的话语权，逐渐成为人们发布信息和表达观点的主要载体。网络舆论成为社会舆情的风向标，并对现实社会产生巨大的影响力和反作用力。近年来多次网络舆论事件的爆发已经清楚的表明，深入研究网络舆论演化机制，有效预测和引导网络舆论迫在眉睫！网络舆论演化是涉及信息科学、网络科学、人类行为学、社会心理学和传播学等诸多领域的复杂问题。面向网络舆论预测和应急响应的需求，本文研究在线社会网络中舆论演化分析的关键技术，揭示网络舆论的形成机制和发展规律。论文的主要工作和创新包括以下几个方面：（1）系统地研究了网络舆论的概念及其演化过程，从方法论的层面探讨了网络舆论演化分析的思路。阐明了网络舆论主体、客体和本体的内涵，剖析了网络舆论的信息属性、社会属性和行为属性，界定了网络舆论的概念；分析了网络舆论的演化过程，论述了网络舆论演化过程中舆论客体、舆论主体所持观点及其状态关系的演化，分别对应于网络话题、网络成员观点以及成员通过相互联系形成的在线社会网络的演化，在此基础上，提出了在线社会网络中舆论演化分析的关键技术问题，建立网络舆论演化分析的技术体系，为后续研究提供理论指导；（2）针对网络舆情信息的海量实时特点和网络舆论处置的应急响应需求，提出了在线话题演化分析框架和方法。在线话题演化分析框架包括子话题发现和关联分析两个部分：子话题发现在线抽取网络信息中隐含的话题片断，关联分析根据子话题间的相互关联组成话题，通过子话题内容和强度的变化描述话题演化。根据上述框架，提出了基于LDA模型的子话题发现方法，定义了子话题产生、消亡、继承、分裂和合并五种演化类型，提出了基于相对熵的子话题时序-内容二维关联分析方法，根据子话题语义相似度和时序关系建立子话题间的关联。基于真实网络新闻和论坛帖子的话题演化分析实验表明，本文提出的在线话题演化分析方法能够有效检测网络话题内容和强度的演化；（3）针对舆论动力学建模的需要，对在线社会网络的结构、动态演化以及网络成员的行为特性进行了深入的实证研究。采用加权有向网络针对某高校大型BBS论坛建立了基于兴趣的在线社会网络（以下简称BBS兴趣网络），对网络成员及其状态关系进行建模。基于复杂网络理论分析了BBS兴趣网络的拓扑结构，实证分析显示，BBS兴趣网络同时具有小世界和无标度特性，主要统计度量指标普遍呈现幂律分布，表明网络成员及其相互联系具有广泛的异质性，并且较其他在线社会网络呈现更密切的成员联系和更强的异质性；进而研究了BBS兴趣网络的动态演化特性，发现BBS兴趣网络的增长具有非平稳特性，网络成员及其相互联系的增长是不均匀的，不同于传统复杂网络理论模型中节点等时间间隔到达、服从均匀分布和连边增长为常数的假设；进一步基于人类动力学方法研究了在线社会网络成员的行为模式，研究表明，网络成员的交互行为具有高度的不均匀性和差异性，呈现长时间的静默与短期高活跃状态交替出现的特性，根据其行为特性，网络成员形成了金字塔形的层级结构，位于层级顶端的成员具有较大的影响力，与其他网络成员存在广泛且密切的联系。上述研究成果为在线社会网络中的舆论动力学提供了实证基础；（4）提出了基于社会影响的离散状态舆论动力学模型，建模网络成员在自我肯定和社会影响双重因素共同作用下如何通过观点改变最终导致网络舆论的产生，定义坚持度和社会影响定量描述上述因素对网络成员观点的影响，基于平均场方法的解析分析表明，基于成员影响力加权占优势的观点是决定舆论演化最终状态的关键因素；在此基础上，研究了本文模型在实证得到的真实BBS兴趣网络中的演化规律，并与主要的复杂网络理论模型进行了对比分析。研究表明，影响力大的网络成员在舆论演化过程中发挥关键作用，其他成员易受该类成员的影响，并倾向于接受其所持的观点。另一方面，网络成员坚持度延缓了一致观点的达成，并且当坚持度大于临界点时导致舆论最终一致态和共存态的相变。不同网络拓扑的对比研究发现，异质网络与同质网络中的舆论演化规律具有显著差异，且异质性强弱影响网络舆论演化的最终状态。无标度网络因其异质性与BBS兴趣网络中的舆论演化规律存在相似之处，而小世界网络和随机网络同属同质网络，舆论演化规律相似。同时，观点初始分布对异质网络中的舆论演化最终状态具有很大影响，而同质网络中舆论演化受观点初始分布影响不大。综上所述，本文深入分析了网络舆论的概念及其演化过程，提出了在线社会网络中舆论演化分析的技术和方法，并通过实证分析和理论建模相结合的方法，验证了本文提出的理论和方法的有效性。本文的研究丰富和发展了网络舆论及其演化的内涵和研究方法，为揭示网络舆论演化的内在机制和本质规律提供了有意的探索，将在网络舆论预测和应急响应等领域发挥重要的作用。更多还原

【Abstract】 In the last decade, the coming-together of technological networks and socialnetworks boosts social interactions through online spaces, which enableinformation exchange and online interactions autonomously with no centralcontrol in a multi-user environment, and exert non-negligible influence on publicopinions. With the increasing importance of network public opinions inunderstanding consensus formation and evolution through online socialnetworks, many efforts have focused recently on evolution analysis of networkpublic opinions. However, due to the unique complexity, the mechanism ofnetwork public opinion evolution on online social networks is far from clearlydefined, although many efforts are contributed. Analysis and modeling ofnetwork public opinion evolution is still an important open problem despite theattentions it has attracted recently.(1) We study concepts and methods of evolution analysis for public opinionsin online social networks. Concepts and processes of the formation of networkpublic opinions are defined firstly. We argue that network public opinionevolution evolves information, social structure and member behaviors, whichdeserves complete and comprehensive study as the foundation of opinionevolution on online social networks. Key techniques implementing evolutionanalysis of network public opinions are further illustrated within the framework.(2) We investigate methods of topic evolution analysis as a part of evolutionof network public opinion. Based on the hierarchy of topics consisting ofcorrelated sub-topics, a framework is proposed including detection andcorrelation analysis of sub-topics. The latent semantics of textual data ismodeled using Latent Dirichlet Allocation (LDA). With consideration of timeinformation, text streams are partitioned into slices, and topic evolution model isproposed, where history topic models provide prior knowledge for topic detectionin the current time-slice. Furthermore, a topic evolution algorithm based on LDAis presented with Bayesian model selection for the appropriate topic numbersand parameters estimation via Gibbs sampling. Furthermore, we define types ofcorrelations of sub-topics, and a method based on relative entropy is proposedto organize correlated sub-topics. We experimentally verify that the method iseffective and efficient for detecting topic evolution of network news and BBSposts.(3) We empirically study the structure and evolution of online socialnetworks using a large scale data set from a forum of BBS. Among various types of online social networks, Bulletin Board System (BBS) is one of the mostpopular to allow people sharing common interests to discuss thoughts or ideason topics. By modeling members of the BBS forum as nodes and theirinteractions as links, we treat the BBS network as a directed graph withconsideration of the closeness of interactions and uncover characteristics of theBBS network, which are fundamental to opinion dynamics on online socialnetworks.More specifically, the mechanisms of growth of nodes and links areinvestigated, indicating mechanisms quite different from existing models.Another important observation is the bilateral scale-free power-law distributionsof in-degree and out-degree, which exhibit significant positive correlation. Thenetwork, on the other hand, shows the “small world phenomenon” with highweighted clustering coefficient and small average shortest path. Furtheranalyses on the dependencies of average strength of nodes as well as averageweighted clustering coefficient on degree confirm the correlations betweenweighted properties and the network topology. The hierarchy of members isproposed, indicating the heterogeneity of member influence. The quantitativeanalysis of member behavior presents power-law distribution of interevent timebetween two consecutive postings in BBS.(4) We discuss opinion dynamics on the BBS network by taking account ofboth social influence and self-affirmation, which addresses state transitions ofactors at the microscopic level, and leads to rich dynamic behaviors at themacroscopic level.Both social influence and individual diversity play important roles in opiniondynamics. Social influence is decisive for individual adoption, which is recentlyverified in online social networks. The more neighbors take an opinion, the morepossibly an actor is convinced to adopt it. On the other hand, online socialnetworks consist of actors of different psychological types and social interactions,which exhibit heterogeneous self-affirmation. At each time step, an actor ischosen to update its opinion according to the interplay of social influence and itspersistence in its current opinion, where each actor is assigned a weightproportional to the power of its strength for its persistence.We investigate the configurations of reaching the final consensuses, andfind that the advantage of weighted fraction, instead of the population, of oneopinion over the other one leads to the consensus. Given a set of typically initialfractions of opinion+1and opinion-1, the consensus converges towards opinion+1and-1, respectively, when the highest-strength or the lowest-strength actorshold opinion+1. Starting from totally random initial distributions, the opinionleading to the consensus features an advantage of the initially weighted fraction over the other, which also holds in the case of equally random distributions oftwo opinions. That is, whether an opinion denominates depends on the initiallyweighted fraction of it. This indicates that high-strength actors play an essentialrole in opinion formation with strong social influence as well as high persistence.Further investigations show that individual diversity slows down the orderingprocess of consensus. Our study provides deep insights into the role of socialinfluence and individual diversity on opinion formation in online social networks.Comparison study shows that opinion evolutions on heterogeneousnetworks and homogeneous networks show dramatic differences. Morespecifically, opinion evolutions on BBS network and scale-free network indicatesimilar characteristics as heterogeneous networks. In contrast, evolutions onsmall-world network and random network are different from those on the aboveheterogeneous networks. Due to the heterogeneity, opinion distributions on BBSnetwork and scale-free network matter.更多还原