在线社会网络中信息扩散研究

发布时间：2018-03-26 19:39

  本文选题：社会网络　切入点：信息扩散　出处：《哈尔滨工业大学》2014年博士论文

【摘要】：Facebook、Twitter等社交类网站的迅猛发展,预示着社会媒体(Social Media)已成为当今网络技术发展的热点和趋势。社会媒体中的用户可以建立各种关系(关注、好友等),从而产生了各种不同的虚拟的在线社会网络。网络中的用户不仅可以发布信息,同时还可以通过共享、转发等行为来传播信息。因此,在线社会网络支撑着信息的发布和扩散。在线社会网络中信息扩散研究可以帮助网络用户获取有用信息、帮助企业推广产品、帮助政府调控舆情,应用价值巨大。本文以真实的在线社会网络数据和信息扩散数据为研究对象,构建了在线社会网络中信息扩散研究的整体框架,并针对研究框架中的用户兴趣描述、信息扩散模型、信息扩散最大化问题、信息扩散和用户推荐相结合等问题展开了研究。本文的研究内容主要包括以下四个部分:传统的信息检索研究中,通常使用词向量来描述用户的兴趣,每个词的权重使用TF-IDF方法来计算。社会化媒体中存在用户、资源和标签这样的三元关系数据,而传统的词向量模型无法充分使用上述三元关系来准确描述用户兴趣,而且词向量方法还存在一词多语义问题。为解决上述问题,本文提出了标签网络模型来描述用户兴趣。在标签网络中,节点代表标签,边代表标签之间的关系。节点和边都是有权重的,代表用户的兴趣度和兴趣间的关联强度。特别的,本文还提出了一种改进的TF-IDF方法来计算标签权重。在Movie Lens和Cite ULike数据集上的实验结果证实了文中提出方法的有效性。信息扩散预测模型可以应用在舆情预警和爆炸性信息识别等方面,具有重要研究意义和应用价值。当前的信息扩散预测模型大多存在两方面问题:一是不具有时间相关的信息扩散预测能力,二是模型训练大都需要耗费较多的时间。为解决这些问题,本文提出了一种新颖的信息扩散预测模型(GT模型)。不同于过去的信息扩散预测模型,在GT模型中,网络中的节点不再被动的受到邻居的影响而执行行为,而是被视为自治的、智能的、理智的个体。用户会计算不同选择下的利益,从而做出理智选择。该模型中引入了时间相关的用户利益,使得GT模型具有了预测信息扩散进程时间动态性的能力。文中创新性的提出了结合全局影响力和社会影响力来计算用户利益的方法。在新浪微博和Flickr数据集上的实验结果验证了文中所提出模型在预测信息扩散时间动态性方面的有效性。当前信息扩散最大化研究基本上都是在无标注社会网络中展开的,这种网络只包含朋友或者信任这类正向关系。然而,信息扩散最大化问题在标注社会网络中的研究仍然是一个有挑战性的并且被忽视的问题。信息扩散最大化研究如果不区分网络用户间关系的极性,将标注社会网络粗略的视为无标注网络,那么用户的正影响力和负影响力都会被误认为正影响力。为解决该问题,本文将信息扩散最大化问题拓展到标注社会网络中,提出了极性相关的信息扩散最大(PRIM)问题和极性相关的独立级联模型,并提出了使用贪心算法来解决该问题。在两个标注社会网络数据集中(Epinions和Slashdot)的实验结果表明,文中提出的方法在解决PRIM问题时要优于未考虑关系极性的贪心算法和其他启发式方法。社会网络主要有两个功能:社会交互和信息扩散。用户推荐研究基于用户的偏好和网络结构帮助用户找到合适的朋友,这就增强了社会网络的社会交互功能。与此同时,用户推荐会促进社会网络中产生新的链接关系,从而加快网络的进化并改变网络结构,而这会直接影响信息扩散,大多数用户推荐方法忽视了这一点。为解决上述问题,文中提出了用户扩散度的概念和计算方法,用户扩散度可以用来对传统推荐算法得到推荐结果进行重排序,从而使得推荐算法可以促进信息扩散。在Email数据和Amazon数据上的实验结果证实了文中所提出的用户扩散度的有效性。此外,本文还提出了可以配合用户扩散度使用的基于超图的用户推荐算法,在新浪微博数据集上的结果表明该方法在推荐指标上要优于过去的方法。
[Abstract]:The rapid development of Facebook, Twitter and other social networking sites, indicates that social media (Social Media) has become a hot spot and trend of the development of network technology. In social media users can build relationships (attention, friends), resulting in a variety of online social network virtual network users only. You can release information, but also through sharing, forwarding and other acts to spread information. Therefore, the online social network to support the dissemination of information and information diffusion. Diffusion research in online social networks can help Internet users to obtain useful information, help enterprises to promote their products, help the government regulation of public opinion, the huge application value. Based on social network data and information online diffusion of real data as the research object, constructs the framework of information diffusion in the online social network, and according to the research framework of the The user interest description, information diffusion model, the diffusion of information maximization problem, information diffusion and user recommendation combination are researched. The main content of this paper includes the following four parts: the study of traditional information retrieval, usually use the word vector to describe the user's interest, the weight of each word using the TF-IDF method to calculate there are users of social media data, three yuan of such resources and tags, and word vector of the traditional model can not make full use of the three yuan to describe user interest, and there are many methods of word vector semantic word problem. To solve the above problems, in this paper the tag network model to describe user interest in the label. In the network, the nodes represent the relationship between the edges represent labels, tags. Nodes and edges are weighted, on behalf of the user of interest and interest relation with the strength. Otherwise, this paper also proposes an improved TF-IDF method to calculate the weights. In the Movie Lens label and Cite ULike data sets. The experimental results confirm the effectiveness of the proposed method in this paper. The information diffusion prediction model can be used in public opinion warning and explosive information recognition. It has important research significance and Application value. The information diffusion prediction model mostly has two problems: one is to do not have the time related information diffusion prediction ability, two is the model training mostly takes more time. To solve these problems, this paper proposes a novel information diffusion model (GT model). Different from the past information diffusion model and in the GT model, the nodes in the network are no longer passive neighbors influence the execution behavior, but is regarded as autonomous, intelligent, rational individuals. The users of the accounting calculation Under the choice of interests, to make rational choice. The time related to the interests of users is introduced in the model, the GT model has the ability to predict the information diffusion process of dynamic time. In the paper, the paper proposed a method to calculate the user interest combined with global influence and social influence. In Sina, micro-blog and Flickr data set the experimental results verify the effectiveness of the model in terms of time dynamic information diffusion prediction proposed in this paper. The current research of information diffusion maximization are basically in no annotation of social network, this network containing only friends or trust this kind of positive relationship. However, the information diffusion maximization problem in marking study on social networks is still a challenging and neglected problem. If the maximum information diffusion study does not distinguish between polar network relationship between users, will mark agency The network will roughly as unlabeled network, then positive influence and negative influence of users will be mistaken for positive influence. In order to solve this problem, this paper will expand to the information diffusion problem of maximizing annotation in social networks, the polar correlation information diffusion (PRIM) and polar correlation independent cascade model, and proposed using the greedy algorithm to solve the problem. In two marked social network data (Epinions and Slashdot). The experimental results show that the proposed method is better than that without considering the relationship between the polarity of the greedy algorithm and other heuristic methods in solving the PRIM problem. Social network has two main functions social interaction and information diffusion. Research on user preferences and network structure to help users find the right friend recommendation based on user, which enhances the social interaction social network. At the same time, with the User recommendation will promote the new generation links in social networks, thus speeding up the evolution of the network and change the network structure, which will directly affect the spread of information, most users recommended methods ignore this. To solve the above problems, this paper puts forward the concept and calculation method of user diffusion degree, the user can spread for the traditional recommendation algorithm for reordering results is recommended, so that the recommendation algorithm can promote the diffusion of information in Email data and Amazon data. The experimental results confirm the validity of the user diffusion is proposed in this paper. In addition, this paper also put forward the user can cooperate with diffusion using recommendation algorithm based on user hypergraph in Sina, micro-blog data sets. The results show that this method is better than the past method in the recommended index.

【学位授予单位】：哈尔滨工业大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP393.092
，

本文编号：1669338