Research On Rumor Detection By Fusing Propagation Network

With the rapid development of the Internet,social media has become the principal place for people to exchange and obtain information.Users can freely publish news and share content on social media,making social media the primary platform for rumor publishing and propagation.Due to the limitation of professional level and cognitive ability,it is difficult for ordinary people to identify rumors from much online information,making it easy to be confused and even become a transit station for rumors.The spread of rumors will cause significant losses to individuals and affect the stability of the whole society.How to automatically identify rumors quickly and accurately has become a research hotspot of researchers at home and abroad.In the current research on rumor detection at home and abroad,many methods are based on the message text.However,on social media,rumor makers often imitate the writing style of the real message to deliberately deceive the reader.The rumor text has strong confusion,which makes it difficult to detect with the text alone.When a message is published on social media,users will comment or forward it,which constitutes a message propagation network.This network contains potential information that is helpful for judging the authenticity of the message.Comments are usually the most direct views and opinions of the public on the original text,and the commentary data is usually on the same web page as the original text,which makes it easy to obtain data.Fully utilizing the original text and the information inside the comment can help improve the performance of rumor detection.Forwarding,like comments,contains public opinions and attitudes,and is larger,but the data quality is uneven and has noise.It is very important for rumor detection to make efficient use of the original text and forwarding network,and to avoid noise interference as much as possible.From the perspective of message propagation,this paper proposes the following two models for comment and forwarding:(1)A rumor detection model Post Com2 DR based on original text and comments.To fully use the original text and comment information,the reply relationship between the original text and comment is modeled as a reply structure diagram.The two-layer graph convolution neural network and self-attention mechanism capture the reply structure information between the original text and comment.Then,the mutual attention mechanism between the original text and comments is introduced to focus on the fusion of information to obtain the global feature representation of the original text and comments.At the same time,to capture the topic changes in the comment sequence,the text convolution neural network is introduced to obtain the local feature representation of the comment sequence.Finally,the global feature representation and local feature representation are fused for rumor detection.(2)A rumor detection model GUCNH based on original text and forwarding network.The message forwarding network on social media platforms is extensive,and the information quality is uneven.To effectively use the information in the forwarding network,firstly,the forwarding graph is constructed according to the forwarding relationship between messages,and a graph convolution neural network module with fusion gating is proposed.The graph convolution neural network is used to capture the structural information in the forwarding graph,and the fusion gating is used to select and combine the information before and after processing through the graph convolution network to enhance the node representation.At the same time,since the association between messages is not limited to the nodes with a direct forwarding relationship,a multi-head self-attention mechanism is introduced to capture the potential impact between any node.Finally,considering the importance of the original information,a gating mechanism is established to selectively enhance the original information,which can effectively prevent the destruction of the original information in the process of information aggregation and improve the performance of rumor detection.The proposed model has been thoroughly tested on multiple Chinese and English datasets such as Rumdect,Weibo-20,Gossip Cop,Twitter,etc.The contents include rumor detection,ablation experiments,early detection,sample analysis,module order research,node pooling methods,etc.And the experimental results are evaluated by indicators such as accuracy,precision,recall,and F1 value.The data results show that the two models proposed in this paper are superior to the current similar mainstream advanced models in rumor detection and early detection.