Research On Microblog Rumor Detection Method Based On Graph Neural Networks

The emergence of microblog has accelerated the speed of information sharing and dissemination,and people’s way of obtaining information is simpler and faster.However,microblog has few restrictions and insufficient supervision,resulting in the emergence and spread of rumors,which has a serious negative impact on the society.With the rapid development of microblog and the increasing number of users,the problems caused by rumors in microblog are getting more and more serious.It is a difficult test to accurately find rumors from a large amount of information in microblog.Traditional microblog rumor detection methods mainly rely on artificial features and machine learning,but the artificial features are not accurate enough to describe the microblog rumor text,so the method of deep learning is introduced.Convolutional Neural Networks(CNN)and other deep learning methods can automatically extract various deep features of rumors,but there are defects that they mainly focus on local features and do not extract enough features between discontinuous words.Therefore,this paper will study the rumor detection method based on Graph Neural Networks(GNN).The main contributions are as follows:(1)A microblog rumor detection method based on Graph Convolutional Networks(GCN)is proposed,which includes three modules: microblog event graph building,node information updating and microblog category recognition.First of all,in order to solve the problem of not making full use of the role of the same word in different microblog events caused by the joint establishment of the whole graph of all microblog events and the problem that the test data needs to appear in the model training,the graph is built separately for each microblog event in the microblog event graph building module.Secondly,in order to solve the problem of insufficient use of context semantics when only using Graph Convolutional Networks for vector updating,this paper makes use of the advantages of Graph Convolutional Networks and Gate Recurrent Unit(GRU)in dealing with continuous and discontinuous text in the node information updating module,and proposes a node vector information propagation model based on Graph Convolutional Networks and Gate Recurrent Unit,making the updated node vector contain more information.Finally,in the microblog category recognition module,this paper proposes a multiple pooling method based on attention mechanism to fuse the max-pooling,average-pooling and global-pooling to obtain the final graph level vector,which effectively integrates the feature information of different node states and realizes the detection of rumors through the loss function.(2)A microblog rumor detection method integrating word semantic relevance and word emotion is proposed,including word relevance graph building,node information aggregation layer and rumor identification layer.Firstly,we use word relevance to build a separate graph for each microblog event,and make effective use of the semantic information between words by building word co-occurrence graph and word semantic correlation graph.Then,in the node information aggregation layer,through the word emotion score table,the nodes are introduced to give emotion weight,make full use of word emotion,and realize the organic integration of word emotion and Graph Neural Networks.Finally,the rumor identification layer is used to identify the rumor.(3)In order to improve the efficiency of microblog rumor detection,this paper explores the impact of microblog comment time on the detection results,Firstly,the time stamp of microblog data is used to obtain microblog comments in different time periods;Then the experiment is carried out by this paper’s method;Finally,by comparing the experimental results,the best comment utilization time threshold for model training is obtained,which makes full use of comment information,reduces the waste of resources,and realizes the early detection of rumor.Compared with other models,the experimental results show that the performance of this paper’s method is better than the typical advanced rumor detection methods on the given data set,which proves the effectiveness of the method in the field of microblog rumor detection.At the same time,the results of ablation experiments also prove that the work of word semantic correlation graph,the fusion of Graph Convolutional Network and Gate Recurrent Unit,the fusion of multiple pooling methods by attention mechanism,and the introduction of word emotion scores for nodes do have an improved effect on rumor detection.