Credibility Analysis Method For User Generated Content In Social Network

Nowadays,we are connected more closely by the emergence of social networks which play an increasingly important role in our common lives.In the network,the number of users is growing rapidly day by day,along with massive user-generated content,which known as UGC.However,due to the feature of openness and rapid propagation,the social network is filled with various of false information,so that users feel great confused when they make judgment about the authenticity of information,moreover,it will even become a direct result of the harmful public opinion,so the studies on the credibility evaluation of user-generated content have become the focus of attention by many experts and scholars in different domains.Especially in the microblogging platform and electronic business platform,which are thought to account for most of the data traffic in the network,it is particularly important to identify credible information on the two platforms.Towards to these two application scenarios,this paper carries on research on the methods of evaluating the credibility of UCG,which mainly contains the following four parts:Firstly,in order to make the distribution of topic more convergent which produced by UGC in a social network,this paper propose Community Awareness–Latent Dirichlet Allocation(CA-LDA).All UGC produced in the whole network can be divided into several communities based on the background and aims,for example,we can create several communities according to users' rating on products,and most of the fivescore comments are positive while one-score comments are opposite.So these comments with the same rating can be regarded as a community.Therefore,we apply the traditional LDA model to different communities and create the CA-LDA model whose topic-word distribution can effectively distinguish the difference of topic distribution between communities and interpret the semantic feature of various UGC accurately.Secondly,in order to reflect the relationship between the credibility and value of UGC,we then propose the comprehensive calculation method of content value that gives quantitative evaluation and analysis.Based on the result of CA-LDA,this paper put forward the integrated quantitative approach by combining with topic entropy,community preference,and the sentiment scores of UGC.This approach can intuitively estimate the quantity of content itself and lay the foundation for evaluating the credibility of UCG.Thirdly,in order to discern credibility differences of UGC between communities effectively,this paper propose the Hierarchical Credibility Network based on CA-LDA(HCNC),this model can effectively assess the credibility of UCG in the social scenario.According to the structure of CALDA model,HCNC model builds the hierarchical structure include document layer,topic layer,and community layer.In terms of the training result of CA-LDA model and the calculation of content value,we set up the connections between the nodes within and outside the layers with some special social attributes,such as repost and thumbs-up,and then give the calculation methods of connection weight and node initial value.Finally,we use the methods of network optimization to simulate the credibility propagation process of HCNC and compute the final credibility value for nodes in every layer by gradient descent algorithm.Fourthly,this paper deploys experiments to validate and analyze the credibility evaluation results of UGC for two actual application scenarios,i.e.,Amazon E-commerce platform and Sina micro-blog platform.We first collect the relevant product comments and blogs about true and fake news events to build assessment dataset.Then we test the performance of some main credibility evolution techniques in this research field,analyze the results and compare with our proposed methods.In sorting experiments,the results show that the performance of our proposed method is remarkably better than all control groups with Normalized Discounted Cumulative Gain(NDGC)index.In classification experiments,the accuracy of our methods is increased by approximately 2%.The research of this paper can help social platforms to filter information,increase the credibility of the decision and also has a certain contribution in evaluation credibility of UCG with the unsupervised method.