Research On Rumor Detection Model Based On Graph Network Structure

With the rapid development of Internet technology and the widespread use of mobile devices,a number of online social media,such as Weibo and Facebook,have gradually become the main carriers of information consumption for the public.The openness and convenience of social media platforms provide people with a space for free expression,but this also indirectly provides an ideal place for the spread of false information.The spread of rumors not only misleads public thinking,but even destabilizes the whole society.Therefore,how to quickly and accurately detect rumors on social media automatically has received extensive attention from domestic and foreign researchers and scholars.Current domestic and foreign research on rumor detection mainly starts from the text content of posts,user information,and the propagation structure of messages,and uses machine learning and deep learning algorithms to model and encode them,and then discriminates rumors.However,posts in social media are short in length and noisy,and there is false information in post content and user attributes,making it difficult to determine the authenticity and credibility of a message,so considering only content information or user information may not be able to extract valid information,thus limiting the detection performance of the model;and the existing work on rumor detection based on the propagation structure either fails to build a structure that matches the actual propagation process at the modeling level and ignores a lot of important information,or the model itself has defects at the coding level and cannot extract more effective features to identify rumors.To address the above problems,this paper proposes two new rumor detection models based on the propagation structure of messages,from both modeling and encoding perspectives.The main contributions of this paper include:(1)Considering that existing rumor detection models based on propagation graphs mainly use graph neural networks to blindly aggregate information of neighboring nodes,which cannot filter nodes unrelated to events and have the problem of oversmoothing,this paper proposes a rumor detection model based on graph pointer network.The model first constructs an undirected propagation graph based on the retweet or comment relationship and initializes it;secondly,a widthfirst search algorithm is used to sample from the source posting nodes to obtain a sequence of nodes;then the local structural features of each node in the propagation graph are extracted using graph convolutional networks;then a pointer network is used to select the most relevant and valuable nodes from the node sequence to obtain a new sequence of nodes,and then one-dimensional convolution and pooling operations are applied to this sequence to capture the high-level semantic features of the event;finally,the initialized features,local structure features and high-level semantic features of the source-posted nodes are combined as the final representations of the event for rumor detection.(2)Considering that most of the existing rumor detection models consider each event as an independent individual and only consider the local information within an event,ignoring the global structural relationship of the event in social media,this paper proposes an event-user heterogeneous graph with time-series information that can reflect both the information within an event and the association between events.In this paper,events and related users are abstracted as two different types of nodes in the network,and the relationship between user nodes and event nodes is established based on the user’s participation in the event(the presence of retweeting or commenting on posts in the event).On the other hand,within each event there is a source post and a series of response posts.In this paper,the response posts are constructed as time sequences based on the time delay of the response posts after the source posting,so that each source post corresponds to a response sequence.Eventually the propagation of messages is constructed as an event-user heterogeneous graph with time-series information.(3)Based on the event-user heterogeneous graph,this paper proposes a sequence-aware heterogeneous graph neural rumor detection model,which aims to mine the complex and diverse features in social media events for identifying rumors.The model models the temporal relationship between response postings within an event by positional encoding and focuses on important response postings using a multi-headed attention mechanism,then fuses source and response postings using an aggregation function of the graph attention mechanism to obtain a local temporal representation of the event;then learns the global structural representation of the event using an element-level attention mechanism based on the interaction between the user and the event;finally,the two feature representations are spliced for rumor classification.(4)In this paper,experiments are conducted on three real scenarios of Twitter15,Twitter16 and Weibo datasets,and the two models proposed in this paper are compared with other models,and accuracy,precision,recall and F1 value are used as evaluation criteria for experimental evaluation.The experiments include rumor detection experiment,ablation experiment,early detection experiment,over-smoothing analysis,and parameter analysis.The experimental results prove that the two models proposed in this paper outperform the current similar mainstream models in rumor detection and early detection.