Research And Implementation Of Social Network Rumor Detection Based On Deep Learning

With the rapid development of Internet,social networks have gradually become one of the important sources for people to obtain news information.While bringing convenience,social networks provide a way for the propagation of harmful rumors as well.Nowadays,rumor detection has become a significant research topic in academia while early detection and accurate detection are two aspects of its concern.The existing rumor detection methods usually use a single model to deal with above challenge,which leads to the following two problems.Firstly,due to the comment dependence of existing rumor detection methods,it is difficult to accurately detect rumors in the early propagation period.Secondly,when the comment data is sufficient to construct the complete way that rumor spread,the existing methods can not fully exploit the the propagation features of rumors,which lead to poor detection performances.In order to solve the problem of comment dependence in the early detection of existing methods,this thesis proposes Early Rumor Detection Based on Garph Attention Network and Variable Auto-Encoder(ERD-GAT-VAE)to mine the deep representation of rumor text and user information for preliminary detection.Based on this,combined with sufficient historical comment data,a multi-task detection method,named Social Network Rumor Detection Based on Bi-Directional Propagation Garph with Multi-task Learning(SNRD-BDPG-MTL),is proposed to detect rumors accurately by effectively mining the deep propagation and wide dispersion features of rumors.Finally,based on the two methods above,a proto-system for social network rumor detection is developed,which provides a complete process of social network rumor detection that contains both early and accurate detection.The main contributions of this thesis are as follows:(1)Aiming at the comment dependence of existing methods in the early period of rumor spreading,this thesis proposes ERD-GAT-VAE based on text and user features to detect rumors.Firstly,without relying on comment data,this thesis proposes Graph Attention over Dependency Tree Model(GADT)to obtain the deep representation vector of rumor text.Secondly,User Credibility Evaluation Model(UCE)is innovatively proposed to obtain the user credibility representation vector.Finally,by combing the deep representation of rumor text and user credibility representation,the early detection result is obtained.Experimental results show that the proposed model can provide more accurate detection results within 12 hours.(2)Focusing on the poor performances of existing methods in the accurate detection of rumors,this thesis proposes a detection method based on Bi-Directional Propagation Garph,named SNRD-BDPG-MTL.First of all,this method proposes Bi-Directional Attention Enhanced Graph Convolutional Network(Bi-AEGCN)to extract the propagation features of rumors.In addition,combined with the statistical features of posts,this method takes the representations of text and user obtained by ERD-GAT-VAE as input while the task of stance detection is introduced as an auxiliary to improve the accuracy of rumor detection.Experimental results show that SNRD-BDPG-MTL achieves good performances in both accuracy and F1 value.(3)Based on the above studies,this thesis designs and implements a proto-system of social network rumor detection.For social network posts submitted by users,the system uses Uniform Content Label(UCL)to normalize the rumor data from different social network platforms and handle early and accurate detection of rumors by using ERD-GAT-VAE and SNRD-BDPGMTL proposed in this thesis,which provides social network users with a complete detection mechanism that includes both early and accurate detection.