Study On The Rumor Stance Classification Based On VAE And GCN

In today's Internet age,Weibo,Twitter,Facebook and other social media platform become the main channel of information initiation and propagation,the rapidity of information dissemination,timeliness and universality has brought great convenience to people's life,but also often become the spread rumors spread(unconfirmed news)when a hotbed of,in turn,affects the security of the network environment,and even cause chaos in the social order.Rumors can be divided into new rumors brought about by breaking news and rumors that have been around and discussed for a long time.Aiming at the characteristics of these two types of rumors,this paper explores the positions of social media users on rumors and designs two different classification models for rumor positions.Firstly,a Variational auto-encoder-long short-term Memory(VAE-LSTM)model was proposed in view of the characteristics of rumors caused by breaking news,such as small number of posts,weak correlation with tweets existing in social media before,unbalanced data distribution and need to be controlled as quickly as possible.Firstly,the model preprocessed the data.After cleaning the tweet data,Word2 Vec model was used to extract the word vectors.Then the word vector is input into VAE for training to get the sequence of depth characteristics in line with the simple probability distribution,and then LSTM network is used to process the vector sequence data.VAE-LSTM model does not need to manually extract or add features,and the training process is relatively simple and efficient,which can meet the timeliness and extensibility of emergency treatment.Moreover,this model can transform data from complex probability distribution to unified simple probability distribution,and then obtain effective features by sampling,which can effectively solve the problem of unbalanced distribution of comment data,and get good results in all evaluation indexes.Then a Graph Rumor Stance Classification(GRSC)model was proposed for the long-existing and long-discussed rumors,which have strong contextual relevance and rich structural information.First initial tweets data modeling into the model Graph data,using the initial data contained in the reply to label and one-to-one tweet ID structure nodes as figure convolution neural network(Graph Convolutional Networks,GCN)input,by drawing attention mechanism structure of tweets Graph embedding,reuse Word2 Vec extract term vector model of text word embedded vector,and combining embedded vector Graph embedding and words,finally using vector obtained by convolution neural network for feature extraction.The GRSC model can extract the context structure information completely by aggregating the attention scores of the firstorder neighborhood of each node.Experimental results also prove the validity of the model.The prototype system of rumor position classification is designed and implemented.Aiming at the above two different types of rumors,based on the data set in this paper,the prototype system of the algorithm in this paper is implemented.The prototype system realizes the functions of tweet information display,tweet classification and data analysis results display.Helps administrators analyze tweets and deal with controversial rumors in a timely manner.This paper contains 35 pictures,9 tables and 83 references.