Research On Identification And Propagation Analysis Of Rumor In Online Social Media

With the continuous development of computer and Internet technologies,information interconnection and human interconnection are highly integrated,and online social media such as social networking sites,blogs and microblogs have been flourishing.Online social media are characterized by fragmented information content,networked information dissemination,and integration of online and offline,which facilitate information exchange while reducing the cost of spreading inaccurate information,thus gradually developing into a breeding ground for rumors.Rumors refer to remarks or statements with content unconfirmed that have been widely spread and cause certain impacts on public opinions.Generally,these remarks or statements fabricate false information by means of creating out of nothing,subjectively fabricating,transplanting and nihilistic history,etc.They would cause the public misunderstanding or triggering their negative emotions,and even increase the possibility of online illegal and criminal activities,which would affect our social stability.Therefore,it is of great practical significance to study the identification methods and dissemination rules of rumors on online social medias,to detect rumors as soon as possible and control the spread of rumors.In this way,we can reduce the harms of rumors to our society or to the public,and create a clear and healthy cyberspace.In this paper,we focus on the study of short text representation of online social medias,rumor identification and rumor propagation,and then some achivements have bee obtained.The details are as follows.(1)A context semantic representation learning model for short textual is designed in view of the characteristics of texts in social medias.The text in online social medias embodies various features such as real-time updating,interactivity,networking and user centrality,etc.,which requires the construction of the representation model with better understanding and better comprehension.First,representations of BERT-based word vectors are deeply studied to obtain a better context representation,and the attention weights for word vectors,sentence vectors and vectors of word locations are introduced and optimized,respectively.Secondly,a dependent syntactic tree based on the attention mechanism is constructed to obtain vector representations with granularity larger than word vectors,to obtain phrase vector representations and semantic representations of texts.Finally,a model for semantic sentiment analysis of text is constructed based on BERT and their syntactic dependency relations.Extensive experiments indicate better information representation capacity of the proposed text semantic representation model.(2)A joint model of stance classification and rumor identification based on dynamic graph representation learning is designed for issues of the complex short textual contexts in social medias.Previous studies considered rumor stance classification and rumor identification as two relatively separate processes,and rumor stance classification has been limited to analyzing each reply post of source rumor information individually,which ignores the context structure information between the replies.In this paper,the rumor stance and rumor identification are thought of as a unified process.And further,the changing positions of the replies are determined by whether the source information is a rumor while different positions of replies could also disprove accordingly whether the source content is a rumor.Based on the above description,a joint model of stance classification and rumor identification based on spatio-temporal dynamic graph representation learning is proposed.First,the issue of rumor identification is converted into a dynamic graph classification problem through the modelling of rumor context by the dynamic graph.Then,the spatial features among the replies are learned by using an improved graph convolution network(GCN)as a spatial relationship model,and their temporal characteristics are learned when utilizing the temporal convolution network(TCN)as a temporal relationship model,respectively.Experiments demonstrate that the proposed spatiotemporal dynamic graph model could simultaneously model the spacial and temporal features in the context of replies,and improve the performance and capacity of the model's rumor identification.(3)A rumor propagation model based on traditional media intervention and rumor dispelling mechanism is designed for the problem that rumor propagation models are usually lack of dispelling and intervening of rumors through the internal nodes of the network.Most current rumor propagation models rely more on the external traditional medias like newspapers,news reports and governmenst regulation and control to conduct the dispelling of rumors.The ignoring of dispelling and intervening of rumors from inside the social networks led to unsatisfactory refuting results because the speed of dispelling is far slower than the speed of rumor propagation.The environments of social medias possess the function of self-purification,however,rarely noticed.That is,groups that will spontaneously release true information do exist even though the social media is flooded with a large number of rumors.Therefore,the spread of rumors will be controlled when the information of rumors and the dispelling of rumors coexist in media enviroment.Consequently,a rumor propagation model based on traditional media intervention and rumor dispelling mechanism is proposed.First,a category of internal nodes named "rumor dispeller" of the network is defined innovatively and they spontaneously take part in the process of the rumor propagation.Except for the rumor immunity nodes,other nodes will convert into "rumor dispeller" with a certain probability after contacting with rumor dispell.Second,the possible influencing factors on rumor spreading are given based on considering the intervention of external traditional means as a factor.Then the equilibrium and stability of the model are analyzed through the theory of mean-field.Finally,simulation experiments were conducted on datasets and the results show that the proposed rumor propagation model,which combines internal rumor dispelling with external traditional interventions,can decrease the speed of rumor propagation tremendously and thus achieve the effect of fast rumor dispelling.