【作者】 汤小虎；

【导师】 刘波；

【作者基本信息】 东南大学 ， 计算机应用技术， 2018， 硕士

【摘要】 近年来社交媒体的迅猛发展拓宽了人们获取信息的渠道,人们可以随时随地的从社交媒体平台上获取各种即时内容。但是社交媒体在给人们获取信息带来巨大便利的同时也促进了虚假信息的传播,给人们造成严重的负面影响。帮助用户识别社交媒体中的虚假信息变得尤为重要。互联网内容可信性的鉴别并不是一个崭新的研究领域,自上个世纪末开始便得到了高度重视。但是相比传统的互联网媒体,社交媒体用户生产内容和开放性的特点使得社交媒体中内容混杂,内容传播的网络结构更加复杂,给内容可信性的判断带来了新的挑战。针对于上述问题,已有的研究在分析社交媒体内容可信性的时候大多忽略了用户的主题特性,缺乏对用户可信性的细粒度考虑,另一方面也忽视了用户从众行为带来的干扰,本文将从这两方面入手展开对社交媒体内容可信性的研究,主要研究内容为:首先,建立社交媒体内容可信性评价模型。在新浪微博公开数据集的基础上,分析社交媒体中用户所体现的主题倾向和从众行为,然后将用户发表或者转发内容的行为视为投票行为,采用生成模型的思想,将用户的主题因素、从众因素以及内容的可信性因素融入到用户投票的生成过程中,利用贝叶斯网络对用户投票的产生进行建模,从而得到社交媒体内容可信性评价模型。其次,求解社交媒体可信性评价模型的参数。一方面,根据吉布斯采样的原理,对模型中隐含变量的联合概率分布和转移概率分布进行推导,从而求得隐含变量的采样规则。另一方面,利用先验分布和似然函数的共轭关系,分别对内容可信性、用户主题分布和从众行为分布的后验概率进行推导,从而得到各个参数的估计值。最后,设计实验验证本文模型的有效性。采用本文提出的社交媒体内容可信性评价模型分析新浪微博公开数据集中微博内容的可信性,验证用户主题分布和从众行为对判断内容可信性的影响,并通过对比实验体现本文提出模型的优势。实验最终结果表明,相对于已有的内容可信性评价模型,本文模型能够更加准确的判断社交媒体内容的可信性。更多还原

【Abstract】 For the past few years,the rapid development of social medium expands the ways of getting information for people.People can get information from social medium anywhere anytime.However,when social medium brings us lots of benefits,it also promotes the spreading of fake information and cause serious side effect.So helping users to discover fake information becomes vital important.Compared with traditional medium,the user generating content and openness of social media makes the information in social media more confounding,the dissemination network more complex and brings new challenges to judge content’s credibility.To solve the problem,existed researches ignored user’s topic factor,lacking consideration of user’s find-grained credibility,what’s more,they also ignored the disturbance of herd mentality.This thesis will consider both sides to evaluate the credibility of content in social medium.The main research work are below:Firstly,create a model to evaluate the credibility of content in social media.Based on the public dataset of Sina microblog,analyze the trend of topic and herd behavior and then treat user’s behavior of posting or reposting content as voting.Use the idea of generative model and Bayesian network to establish a model for the generation of a vote with user’s topic factor,herd mentality factor and the credibility of content.Secondly,solve the model.On one hand,using the idea of Gibbs sampling,infer the joint probability distribution and transition probability distribution of all latent variables to achieve the sampling rules for latent variables.On the other hand,use the relationship between prior distribution and likelihood function to estimate the distribution of content’s credibility,user’s topic and herd behavior.Finally,design experiments to verify our model.Use our model to analyze the credibility of content in the dataset and how topic factor and herd mentality factor influence the judgment of content’s credibility.Verify our model’s effect through comparing to existed models and the result shows that our model is more effective.更多还原