基于特征学习的社交网络用户影响力计算模型研究

发布时间：2018-01-01 23:08

本文关键词：基于特征学习的社交网络用户影响力计算模型研究　出处：《合肥工业大学》2015年硕士论文　论文类型：学位论文

【摘要】：近年来,随着互联网的快速发展与应用,在线社交网络迅速流行。基于特定用户的社会影响力在社交网络中的多种应用(如病毒式网络营销或推荐系统)中扮演着重要的角色,因而对于用户社会影响力的分析计算的研究变得迫切需要。目前虽然对于社会影响力的研究层出不穷,但现有方法通常通过对社交关系的统计来隐性表示用户之间的影响力,缺乏深度分析。与此不同,本文中的方法对用户之间影响力的概率进行显性建模,从影响力的产生和传播本质上对其进行分析。本文中所提出的方法尝试从根源上解答为何用户会受到影响以及是否可以通过分析网络上现有的影响力现象得出新的推断这两个问题。通过这种方法,我们不仅能够对现有的影响力进行描述,而且能够对未知用户和未知话题的影响力进行预测。本文的主要工作分为以下三点：1.从微观和宏观角度对社交网络中用户影响力进行全面的定义。针对现有的两种描述型和预测型影响力分析研究工作进行概括介绍并分析其存在的问题。2.提出一种新型的基于内容和社交关系的影响力特征学习方法,通过对微观影响力概率进行显性建模,为影响力构建一种普适性的特征表示。并通过对异步分布式随机梯度下降算法的研究,将本文所提方法有效运用在真实世界的大型社交网络中,学习具有预测特性的影响力特征。3.通过在一个含有五千多万用户和八亿多微博的真实社交网络数据集上的大量实验表明,不管是在微观还是宏观层面的影响力计算和预测方面,本文中提出的方法对比其他现有的方法都有着显著的提高。
[Abstract]:In recent years, with the rapid development and application of the Internet, the rapid popularity of online social networks. Based on the specific user's social influence in social network applications (such as viral network marketing or recommendation system) plays an important role, so study on social influence analysis of user computing has become an urgent need for research. Although the social influence but the existing methods usually emerge in an endless stream, by analyzing the social relationships between users said to recessive influence, lack of depth analysis. By contrast, the probability of the method in this paper on the influence of the dominant user modeling, analysis from the influence of the generation and propagation of the essence of the method proposed in this paper. Try to answer from the root of why users will be affected and whether you can through the analysis of the phenomenon of the network that influence The new inference of the two questions. Through this method, we can not only describe the existing influence, but also to unknown users and unknown topic influence prediction. The main work of this paper is divided into the following three points: 1. from the micro and macro perspective on the social network user influence comprehensive definition for the two. A description of the existing and predicted impact analysis to introduce and summarize the research work and analyzes the problems of.2. put forward a new learning content and influence characteristics of social relationships based on a method by explicit modeling of the micro influence probability, expressed as the influence to build a universal feature. And by descent algorithm research on asynchronous distributed stochastic gradient, the proposed method effectively using large social networks in the real world, learning has the characteristics of prediction Influence characteristics through a.3. containing a large number of experiment about fifty million users and more than 800000000 of micro-blog's real social network data sets show that both calculation and prediction in micro or macro level influence, comparing the method proposed in this paper to other existing methods have improved significantly.

【学位授予单位】：合肥工业大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：G206

【相似文献】