基于网络化数据分析的社会计算关键问题研究

发布时间：2018-05-18 06:01

本文选题：社会化网络 + 网络群体行为　；参考：《北京邮电大学》2014年博士论文

【摘要】：随着信息技术的发展,尤其是云计算、物联网、社会化网络以及信息获取技术的进步,人类逐渐步入大数据时代。数据规模的飞速增长,促使了结合计算科学和社会科学的交叉学科：社会计算的诞生。社会计算旨在利用计算机技术和社科理论搭建虚拟网络与现实社会之间的桥梁,通过对网络化数据的分析揭示网络群体的交互规律,帮助人们认识和研究社会科学的各种问题。本文围绕社会计算中的若干关键问题,从社会化网络的群体互动规律、网络虚拟社区发现和群体舆论的聚集机理三个方面入手,研究相关算法和模型,其主要创新性工作概括如下： 1.针对社会化网络中的即时通信社区群体特点,分析微博社区与即时通信社区在传播范围、隐私性、实时性、用户参与度和会话特性共五方面的异同。通过对即时通信社区和微博社区的用户行为进行分析,提出群体互动指数Sj-inf和群体惰性Iinert两个驱动用户行为的动力学指标,并在此基础上,提出一种基于兴趣与群体互动驱动的行为模型ICHM。模型考虑了群体互动指数、群体惰性和个体兴趣三个影响用户行为的动力学因素。实验结果表明,ICHM生成的群体信息发布行为的时间间隔服从单一指数的幂律分布。通过参数调整可以得到与真实数据相近的幂律分布特性,与实际数据表现的动力学特征相吻合。能够为即时通信社区的群体行为特征提供合理有效的解释。 2.针对传统人类动力学实证结果中出现的指数截断特性,结合即时通信社区个体之间以会话为中心的交互特性,在ICHM模型的基础上,进一步提出基于会话驱动的用户行为模型SICHM。模型基于个体会话交互原则,引入会话转移概率Ptrans和会话退出概率Pcancel约束个体的信息发布行为。实验结果表明,在会话驱动的交互行为中,个体信息发布的时间间隔服从带有指数截断的幂律分布,用户会话转移概率Ptrans在一定区间内会影响两段幂率的幂指数和截断位,当Ptrans偏离该区间范围,个体行为可以用单一幂指数的幂函数刻画。通过参数调整,模型可以生成与真实数据集一致的带有指数截断的幂律分布,表明会话驱动的特性是导致人类行为的幂律特性产生指数截断的原因之 3.针对传统社区发现算法的效率问题,提出一种适用于有向网络社区发现的网络稀疏化算法。算法基于邻居节点之间的共引、传递和耦合三种关系计算其归一化相似度,并引入Minwise哈希函数提高相似度计算的效率。在此基础上,提出基于局部相似度的网络稀疏化算法LSDN。算法充分考虑网络的局部特性,使得网络在稀疏前后的入度和出度均服从幂律分布,从而保证网络稀疏前后的整体分布特性。实验结果表明,所提出的网络稀疏算法可以有效地对有向网络进行简化,能够在不改变网络整体特性的同时保证社区发现的准确度,从而提升社区发现的效率,解决了传统社区发现算法在可扩展性方面考虑不足和效率不高的问题。 4.针对传统舆论演化模型忽略群体一致性压力,缺乏个体从众效应下决策行为研究的问题,分析网络群体舆论演化的驱动力,提出—种基于决策偏移的舆论演化动力学模型DO2M。模型基于社会心理学的从众效应,引入群体一致性压力和个期望牵引力,建立依从、趋同和内化三种不同节点的状态转移策略和观点演化策略。同时,基于决策偏移思想在经典有限信任模型HK模型上进行改进,构建HK-DO2M模型。实验结果表明,提出的两个模型均能够有效模拟群体舆论演化的收敛和分化,与经典有限信任模型相比,模型将个体的期望观点和实际观点分离,其观点偏移的特性更加符合社会网络中群体舆论演进和个体交互的行为特征,揭示了群体层面的观点演化的内在规律,为大数据时代分析现实舆论形成的内在机理提供理论模型和参考。
[Abstract]:With the development of information technology, especially the progress of cloud computing, Internet of things, social networks and information acquisition technology, human beings have gradually stepped into the era of big data. The rapid growth of data scale has prompted the interdisciplinary science and social sciences to be combined: the birth of social computing.
Social computing aims to build a bridge between virtual network and real society by using computer technology and social science theory. Through the analysis of network data, it reveals the interaction rules of network groups and helps people to understand and study the various problems of social science. This paper focuses on some key problems in social computing, from the group of social networks. There are three aspects of the law of body interaction, the discovery of network virtual community and the aggregation mechanism of group public opinion, and the related algorithms and models are studied. The main innovative work is as follows:
1. to analyze the similarities and differences between the micro-blog community and the instant communication community in the five aspects of the communication range, privacy, real time, user participation and conversation characteristics in the social network, and analyze the user behavior of the instant communication community and the micro-blog community, and propose the group interaction index Sj-inf and the group laziness. On the basis of the two dynamic indicators that drive user behavior, a behavioral model ICHM. model based on interest and group interaction is proposed, and three dynamic factors affecting user behavior are considered, including group interaction index, group inertia and individual interest. Experimental results show that the group information release behavior generated by ICHM is generated by ICHM. The time interval obeys the power law distribution of the single exponent. The power law distribution similar to the real data can be obtained by the parameter adjustment, which is consistent with the dynamic characteristics of the actual data. It can provide a reasonable and effective explanation for the community behavior characteristics of the instant communication community.
2. on the basis of ICHM model, the SICHM. model of user behavior model based on session driven based on the interaction principle of experience language, and introducing the session transfer probability Ptrans, and the interactive characteristics of the traditional human dynamics empirical results, combined with the conversation centered interaction between the individuals in the instant communication community. The session exit probability Pcancel restricts the information release behavior of the individual. The experimental results show that, in the session driven interaction, the time interval of individual information release obeys exponential truncated power law distribution, and the user session transfer probability Ptrans affects the power exponent and the cutoff of the two power exponents in a certain interval, when Ptrans deviates from that Interval range, individual behavior can be portrayed by a power function of a single power exponent. By adjusting the parameters, the model can generate an exponential truncated power law distribution consistent with the real dataset, indicating that the characteristic of session driven is the reason why the power law characteristic of human behavior is truncated exponentially.
3. in view of the efficiency of the traditional community discovery algorithm, a network sparsity algorithm is proposed for the discovery of a directed network community. The algorithm calculates its normalized similarity based on the three relationships of the co citation, transfer and coupling between the neighbor nodes, and introduces the Minwise hash function to improve the efficiency of the similarity calculation. On this basis, the base is proposed. The local similarity based network sparsation algorithm LSDN. takes full consideration of the local characteristics of the network, which makes the admission and output of the network obey the power law distribution before and after the sparsity, so as to ensure the overall distribution of the network before and after the sparse network. The experimental results show that the proposed network sparse algorithm can effectively simplify the directed network. It can ensure the accuracy of community discovery without changing the overall characteristics of the network, so as to improve the efficiency of community discovery, and solve the problem that the traditional community discovery algorithm is insufficient and inefficient in the aspect of extensibility.
4. the traditional opinion evolution model ignores the group consistency pressure, lacks the problem of the decision behavior research under the individual herd effect, analyzes the driving force of the network group's public opinion evolution, and proposes a public opinion evolution dynamic model DO2M. model based on the decision shift, which is based on the herd effect of the sociopsycho psychology, and introduces the group consistency pressure and the one. We expect the traction force to establish the state transfer strategy and the viewpoint evolution strategy of three different nodes, including compliance, convergence and internalization. At the same time, based on the decision migration idea, the HK-DO2M model is constructed on the classic finite trust model HK model. The experimental results show that the proposed two models can effectively simulate the convergence and convergence of the group opinion evolution. Differentiation, compared with the classic finite trust model, the model separations the individual's expectation from the actual point of view. The characteristic of the view offset is more consistent with the evolution of the public opinion and the behavior characteristics of the individual interaction in the social network, reveals the inherent law of the view evolution of the group level, and is the inner machine for the analysis of the reality public opinion in the large data age. The theory provides a theoretical model and reference.
【学位授予单位】：北京邮电大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP393.01

【参考文献】