基于时序的社交网络因果关系发现

发布时间：2018-05-28 03:41

本文选题：因果网络 + 因果推断　；参考：《广东工业大学》2016年硕士论文

【摘要】：随着社交网络的飞速发展,越来越多的人开始挖掘社交网络潜在的价值,进而推动相关产业的发展,例如微商、微博营销、社交化电商等。在社交网络众多相关研究中,用户影响力对舆论引导、微博营销具有现实意义,是当前研究的难点和热点。现有的研究用户影响力的方法主要基于用户显式声明的好友网络,然而用户显式声明的好友网络往往具有较大的冗余性。具体表现为大量显式声明的好友网络对于用户的影响力没有实质作用。因此,如何基于用户行为数据,挖掘用户行为之间的因果网络是用户影响力评估的关键,具有重要的意义。然而,已有的社交网络因果关系推断方法中存在两个问题：1.无法识别间接因果影响而导致因果网络出现大量冗余边；2.没有充分考虑因果影响滞后长度。本文基于最小描述长度准则(MDL)对上述两个问题进行了统一建模,提出了一种新的模型MCRNC Minimal Causal Network)。在减少因果网络冗余方面,MCRN模型将因果传递熵算法应用于社交网络因果关系发现上,同时结合因果影响滞后长度对因果传递熵进行拓展,上述策略有效剔除了结构中的冗余边,提高算法的精确率；在探索因果影响滞后长度方面,使用MDL作为模型的评分标准,权衡模型的不确性和复杂度,有效降低了模型过拟合的问题。本文通过大量模拟数据集验证MCRN模型的多个评测指标都优于传递熵,因果传递熵等类似算法。通过新浪微博真实数据集的实验发现用户显式声明的好友关系很多不存在因果影响,存在因果关系的用户之间存在互动行为等现象,较好地验证了本模型的有效性。最后,本文以MCRN模型为理论基础,提出一个基于时序的社交网络因果关系发现系统的构建方案,简称MCRN系统,并给出系统架构和系统原型,该系统有助于用户准确直观地分析用户之间的因果关系,并进一步应用于现实生活中的其他领域。
[Abstract]:With the rapid development of social networks, more and more people begin to tap the potential value of social networks, and then promote the development of related industries, such as micro-quotient, Weibo marketing, social e-commerce and so on. In many related researches of social network, user influence has practical significance to guide public opinion and Weibo marketing, which is the difficulty and hot spot of current research. The existing methods to study user's influence are mainly based on the user's explicitly declared friend network, however, the user's explicitly declared friend network is often redundant. It shows that a large number of explicit friend networks have no real impact on users. Therefore, how to mine causal networks between user behaviors based on user behavior data is the key of user impact assessment and has important significance. However, there are two problems in the existing causality inference methods of social networks: 1. The failure to identify indirect causal effects leads to a large number of redundant side effects in causal networks. The lag length of causality is not fully taken into account. This paper presents a unified modeling of the two problems based on the minimum description length criterion (MDL), and proposes a new model, MCRNC Minimal Causal Network. In the aspect of reducing redundancy of causality network, MCRN model applies causality transfer entropy algorithm to the discovery of causality in social network, and extends causality transfer entropy by combining the length of causality influence lag. The above strategy can effectively eliminate redundant edges in the structure. In order to improve the accuracy rate of the algorithm and to explore the delay length of causality, MDL is used as the scoring criterion of the model to weigh the uncertainty and complexity of the model, and the problem of model overfitting is effectively reduced. In this paper, a large number of simulated data sets are used to verify that many evaluation indexes of MCRN model are superior to similar algorithms such as transfer entropy, causal transfer entropy and so on. Through the experiments of the real data set of Sina Weibo, it is found that there is no causality in many of the user's declared friendships, and there are some phenomena such as interactive behavior among the users who have the causal relationship. The validity of this model is well verified. Finally, based on the MCRN model, this paper proposes a time-series based causality discovery system for social networks, called MCRN system, and gives the architecture and prototype of the system. The system can help users analyze the causality between users accurately and intuitively, and further apply it to other fields in real life.
【学位授予单位】：广东工业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP393.09

【相似文献】