Research On Hot Event Prediction Technology Based On Graph Neural Network

With fast information transmission speed,wide range,and high immediacy,social networks attract large numbers of users to share social news and discuss reality events.Social network events are usually the mapping of real events in cyberspace,so it is of great significance and value to study the prediction of hot events in social networks.For example,predicting hot events can help relevant departments control rumor spread,avoid social panic and safeguard public safety.The information released by users on social networks is mainly text content,which has a high correlation with the heat of an event and is easier to obtain than other event-related data.Text is a vital data basis for event prediction.However,texts are of sparse value,difficult to mine potential valuable information,and event prediction is difficult and complex.In recent years,based on text data,graph neural networks have been widely used in various tasks and achieved good results.Graph neural networks can fully explore the interdependence between keywords,highlight valuable information,and solve the problem of sparse data value to a certain extent.In this thesis,the graph neural network is the core technology for social network hot event prediction research.The main work is as follows:(1)A hot event prediction model based on features gated encode graph convolutional network is proposed.For the short text of micro-blog social networks,the existing dynamic graphic convolutional neural network does not take full advantage of temporal characteristics and semantic information.The dynamic graph convolutional networks are used to solve the hot event prediction problem,and its time-series encode module is improved.The feature vectors after the convolution of the graph are compressed,and the feature statistics of each dimension are obtained by global average pooling operation.The summary of key features is generated and used to scale the original features.The experimental datasets are built based on real social networks.The results prove that the proposed model can effectively utilize text information of social networks to improve event prediction performance.(2)A hot event prediction model based on features fusion dynamic graph convolutional network is proposed.Social network texts,time-series numerical information such as the number of an event forwards,can describe event heat from different perspectives,which are interrelated and complementary to each other.A more comprehensive event evolution model can be learned by comprehensively utilizing multiple types of information.Based on the model in Chapter 3,a gated recurrent unit is used to model the temporal correlation of numerical information,and the obtained numerical eigenvectors are fused with the textual eigenvectors to realize the hot event prediction.The comparative experimental results demonstrate that the fusion of multiclass features can effectively improve the accuracy of hot event prediction.(3)A hot event prediction system based on features fusion dynamic graph network is designed and implemented.According to the above research contents,a hot event prediction system is constructed,which is of positive significance for assisting relevant departments to control rumor propagation and avoid social panic.