Research On Information Diffusion Prediction Method Based On Social Graph And Diffusion Grap

With the increasing popularity and widespread use of social networking platforms,social networking platforms have gradually integrated into people’s daily lives,while people are increasingly enjoying the convenience brought by the network platform.Social networking platforms allow people to share and access various news.However,how information is disseminated in the network and what factors influence the direction of information flow have gradually become hot research topics in the field of social network analysis.Most of the current studies either only use the information diffusion sequence or only use the social network among users to make predictions,which makes it difficult to effectively model the complexity of the information diffusion process.To solve this problem,three information diffusion prediction methods are proposed in this paper.These methods focus on studying the multiple effects of social graphs and diffusion graphs on information diffusion prediction,and the main contributions of this paper are as follows:(1)Dynamic heterogeneous graph perception network with time-based mini-batch for information diffusion prediction(DHGPNTM),constructs social graphs and diffusion graphs based on the user’s forwarding time into the network.The proposed graph perception network(GPN)is used to learn the dynamic structural features of users and is then combined with the timing of users’ forwarding information to learn the dynamic preferences of users.The proposed fusion gate mechanism is used to selectively fuse user information with contextual information.Finally,information diffusion prediction is performed.(2)Social information diffusion prediction method based on transformer spatialtemporal neural network(TSTNet),this method forms a whole heterogeneous graph by social graphs and diffusion graphs through GCN to learn the user’s preferences.The structural features and temporal features are then embedded in the position information;because the transformer does not learn the relative position information,the temporal features after embedding the position information are used as the Query set,and the structural features after embedding the position information are used as the Key set and Value set.The Transformer mechanism is then used to obtain the spatial-temporal features of users,and finally,information diffusion prediction is performed.(3)Multi-aware attention model for information diffusion prediction(MAA),this method learns the social features and the diffusion features of users through GCN and the timing of users’ forwarding information through the proposed time-based mini-batch module.Then,the proposed feedforward neural network with position awareness acquires position-aware attention and finally performs information diffusion prediction.