Popularity Prediction Of Sina Weibo Combining Text And Time Series Information Based On Graph Neural Network

Accurate prediction of microblogs popularity can give advance cognition and warning to social hot topics and public opinion,which is widely used in advertising,recommendation system and other fields.At present,a lot of work focuses on the feature of microblog themselves,while ignoring the external factors affecting the popularity of microblog.This paper introduces graph neural network(GCN)to construct the relations between microblogs to consider the external factors of microblogs to a certain extent.How to effectively use the information of microblog to consider the relations between microblogs and predict the future microblog efficiently has become key problems.In view of the above problems,the following two aspects have been carried out.In order to solve the problem of graph construction in graph neural network,this paper proposed dual text graph neural network(DTGNN)model.DTGNN takes the influence relations between similar microblog text contents as external factors,uses text and keywords to construct text graph,and then uses GCN and deep graph Infomax(DGI)to obtain the node representation.At the same time,the time information of microblog itself is used as the main influencing factor to fuse with the node representation to obtain the prediction results.When DTGNN is applied to Microblog PCU dataset,compared with the best model using only its own features,MSE is decreased from 0.214 to 0.148 and MAPE is decreased from 0.197 to 0.194.To further improve the performance,considering the corresponding relations between time information and the problem that GCN can not predict the popularity of future microblog,this paper proposed fusion of text and time subgraph neural network(FTSGNN).FTSGNN adds time graph construction part and data segmentation part on the basis of DTGNN.When constructing the graph,FTSGNN considers not only the relations between microblog texts,but also the relations between microblog time.The data segmentation part focuses on the impact of current public opinion on the popularity of microblog.Compared with DTGNN,the speed of model training is shortened from 1953 seconds to 90 seconds.When the 2-hour and 4-hour time information is known,compared with DTGNN,the MSE of FTSGNN is reduced from 0.229 to 0.226 and 0.265 to 0.247,respectively.This paper integrates the relations between microblog text content and time information with the information of microblog itself,which improved the prediction performance compared with using the feature of microblog itself.At the same time,the subgraph was divided by data segmentation,which solved the problem that GCN can not predict the new microblog to a certain extent.The experimental results proved the effectiveness of DTGNN and FTSGNN models and the influence of external factors on microblog popularity.