Research On Information Popularity Prediction Based On Spatio And Temporal Features

With the rapid development of network technology,Internet has served as the main platform for people to obtain and share information.Compared with traditional word-of-mouth way of information dissemination,information spreading on Internet exhibits faster and wider range.On the one hand,people can access massive amounts of information more conveniently and quickly,they also suffer from the adverse effects of fake news and information cocoon.Predicting the trend of information dissemination on the internet and determining the popularity of information can not only effectively control the negative impact brought by fake news,but also provide theoretical guidance for dissemination of positive information.Therefore,research on prediction of information popularity has great significance and application value for cyberspace security governance.The existing information popularity prediction mainly carries out research from two perspectives of dynamics mechanism and feature modeling,focusing on user,structural and temporal characteristics of information cascade.However,the existing models still have shortcomings in feature extraction and fusion.As such,this paper focuses on spatio-temporal feature extraction and fusion of information dissemination.The main research content and innovation points are as follows:(1)To address the problem that the complex temporal dynamics of the information cascade is difficult to model,a popularity prediction model based on the fusion of spatio-temporal feature,Cas GLT,is proposed,which uses GCN for presentation of structural features of information cascade,and adopts attention mechanism to learn the expression of association relations of nodes.In addition,the model also utilizes LSTM and TCN for more adequate extraction of temporal features,and obtains information cascade representation vectors by fusing structural and temporal features.After that,the vectors are fed into a multilayer perceptron for popularity prediction.Experiments conducted on two publicly available datasets validate the effectiveness of the Cas GLT model,achieving 12% and 6%improvement in prediction accuracy,respectively.(2)To address the problem that dynamic changes in the information cascade structure are difficult to express,a popularity prediction model based on spatio-temporal neighborhood embedding,ADJConv,is proposed.The spatio-temporal neighborhood of each node is extracted based on the adjacency matrix of the information cascade graph.The features of the nodes within the spatio-temporal neighborhood are aggregated using a convolutional neural network,a new iterative convolutional method is proposed to learn deeper spatio-temporal neighborhood features of the nodes.With that,the expression vectors are used to predict the popularity.The experimental results of the ADJConv model on two publicly available datasets outperform the baseline method,demonstrating the effectiveness of the model in the popularity prediction task and achieving 7% and 12% improvement in prediction accuracy,respectively.