| With the rapid development of the Internet and the online social meida,social network related industrial products have been widely used in people’s daily life,such as Wechat,Microblog,QQ,Twitter,and Facebook.At present,these products have become the most commonly used online communication tools for public.Based on these pruducts,the information in social networks owns important characteristics of fast speeding and wide transmission range.Moreover,due to the massive scale of the online social network,the virtual social network can directly affect the real society.In the online social network,unverified information can be easily made and widely spread,netizens who do not know the veracity may believe the unverified information and diffuse these unverified information(the unverified information is defined as rumor),which affects national security and social stability to some certain extent.Therefore,effective rumor analysis tools can recognize rumors in the online social network to reduce the negative impact of emergencies and promote the healthy and harmonious development of society.This thesis analyzes the rumor problem from the perspective of rumor information dissemination detection,prevention and early warning.The contained rumor analysis tasks are divided into rumor detection,rumor stance classification,rumor sources detection,and rumor veracity detection.For these rumor tasks,this thesis studies the key technologies of rumor analysis in social networks based on machine learning methods,which mainly contain the following four aspects:(1)In the rumor detection task,neural networks based on graph structure representation and tree structure representation have their own advantages for rumor detection.To extract the global and local structure features between source and reply in rumor sessions,this thesis constructs the structure expression between sources and their replies for each conversation from the perspective of graph and tree.By constructing a graph representation that mainly focuses on modeling global structure information,an Ensemble Graph Convolutional neural Network with node proportion allocation mechanism(EGCN)is proposed.By constructing a tree representation that mainly focuses on modeling the reply relationship of nodes in the neighborhood,a source reply conversation Tree Convolutional neural Network(TCN)is proposed.The proposed EGCN and TCN are directly used in rumor detection task respectively.The accuracies and F1 scores of the EGCN and TCN models are tested on five events of rumor detection.The experimental results verify the effectiveness of the proposed two models in rumor detection task.(2)For the rumor stance classification scenario of independent news and the rumor stance classification scenario of complete rumor conversations,this paper proposes rumor stacne classification models for the two scenarios by building corresponding attention mechanisms.Firstly,according to the information granularity of the text and the public’s reading and expression habits,a rumor stance classification model based on Stochastic Attention Convolution Nneural Network(SACNN)is proposed for independent news;For the complete conversation,from the perspective of extracting the structural information between the source node and the replies,a rumor stacne classification model based on Attention Tree Convolution Neural Network(ATCNN)is proposed.Comparative experiments and ablation experiments on 8 social network events of the public dataset verify the effectiveness of the above two models in the rumor stance classification task.(3)To construct effective rumor source detection methods for social networks with different structures and different forms of propagation,this thesis transforms the rumor source detection problem into the prediction for source nodes in social networks from the view of pattern recognition.Based on the attention mechanism,a High-order Graph self Attention neural network(HGAT)model for rumor source detection is proposed.The neighborhood structure of nodes with different orders is used to model the propagation information of rumors in social networks,so as to realize the source detection task for social networks with different structures.The HGAT model is verified by different propagation models and real network datasets.Moreover,the experiments also verify that the high order attention mechanism is effective for the source detection task.(4)In the rumor veracity detection task,from the perspective of multi-task learning and structural information modeling,firstly,rumor conversations with different structures are rebuilt as Regularized full M-ary rumor Conversation tree(RC tree)with the same structure and different neighborhood node features;Based on RC trees,a Multi task self Attention Tree Convolution Network(MATCN)based on RC tree is proposed according to the auxiliary effect of stacne features in rumor conversation on the rumor veracity detection task.The proposed MATCN builds a tree self attention mechanism for node classification task(rumor stance classification),and constructs conversation tree convolution and conversation tree pooling method for RC tree classification(rumor veracity detection).Based on the multi task learning framework,the proposed MATCN optimizes the veracity detection results with the help of stacne features.The results of comparative experiments and ablation experiments on the public rumor veracity detection dataset show that the MATCN can improve the effects of rumor veracity detection compared with other neural networks based on graph or tree structure. |
