Misinformation Identification And Propagation Analysis In Social Network

Recently,with the rapid development of smart device and mobile Internet,online social network provides more and more convenient and quick access for people to sharing and obtaining information.However,the quality problem of information concerns everyone at the same time.All kinds of misinformation are spread over the Internet,giving rise to severely negative impact on the public life and social stability.The research on misinformation identification for social media has been promoted,with the intention of using various characters of misinformation,to achieve early detection of misinformation hidden in social media.Presently,existing research mainly focuses on the diffusion process of information over social network,to explore the potential features in the propagation context of misinformation.Most researchers tend to model the diffusion process as the sequence structure,based on the chronological relationship between tweets.But most of the approaches rely on the text of tweet that could be controlled by people easily,which is deficient for misinformation detecting,such as hiring spammers to repost misinformation or deleting the comments.In addition,related works neglect the semantic interaction between tweets,which could model the evolution process of information more thoroughly.The propagation tree has been involved in to model the diffusion process in latest works,most of which using Recursive Neural Networks(Rv NN)to extract the feature of propagation tree.But due to the limitation of Rv NN,it is inefficient for these models to learn the long-term dependency and hidden features.In view of the deficiencies in the existing research,this thesis focuses on the evolution of propagation process of tweet in social media,and designs an effective misinformation identification model based on the propagation tree.The main works are as follows:Firstly,a real dataset including misinformation and real information is collected from Sina Weibo based on Scrapy.We design a customized webpages crawler for the platform to collect both fake and real microblogs,according to the Sina Weibo management platform.After cleaning the raw dataset,the relevant statistics of the dataset are analyzed to verify the effectiveness.Then,an analysis method based on the propagation tree is proposed to represent the dynamic evolution of information,including the propagation tree and the chronological relationship between tweets.According to the shortcomings of related works,that it is difficult to extract the hidden feature of propagation structure and thus cannot describe the dynamic evolution of propagation tree,the method based on binary tree transformation and subtree decomposition is proposed to extract fine-grained structural feature,and the tree sequence is constructed to represent the evolution of propagation tree.On the other hand,for the chronological relationship between tweets,a Co-Attention mechanism is designed to fuse user attribute features and text features.A classification model is constructed and compared with the existing misinformation detection models,to verify the applicability of proposed methods.The experiment results show that the propagation structural feature and the internal correlation between content and user are effective for the identification of misinformation.Furthermore,a misinformation identification model(Tree-TF)is designed based on the interactive relationship in propagation tree.Focusing on the problem of early detection of misinformation,we proposed the Tree-Transformer framework based on the interaction of tweets in propagation process by mapping the structural feature to Transformer model.And then the Masked Multi-head Co-Attention mechanism is designed in the framework to involve the internal correlation between the user and the content.For the first time,the user feature and text are integrated in modelling the propagation tree,and the misinformation identification model is proposed based on Tree-Transformer.Finally,the dataset Mis Infdect and the public dataset Rumdect are involved to design compare experiments with the existing approaches.Compared with the latest research,such as the works published on Computer & Security 2019,AAAI 2020 and so on,the Tree-TF model proposed in this thesis achieved the best result in both datasets.The accuracy of Tree-TF in public dataset Rumdect has increased by 1.9% and1.1% respectively.For early identification,Tree-TF performed best even at the beginning of propagation.We designed the misinformation identification system in the end.The research of misinformation identification based on the analysis of information propagation process can help to extract the features of information diffusion.An effective misinformation identification model is proposed,which can help related websites to improve the performance of misinformation identification,and reduce the negative effects caused by the spread of misinformation.