Predictive Modeling of Information Propagation Based on Modulated Position Coded Embedded Transformer and Graph Attention Networks

With the rise of social media and online communities, information dissemination and content forwarding among users have become one of the topics of great interest in the field of social network analysis. Traditional information dissemination prediction models are usually based on simple statistical methods or local network structures, ignoring the heterogeneity of social networks and the global dynamic characteristics of information dissemination. These models often fail to capture the complex interactions and temporal dependencies in the information propagation process, resulting in limited prediction accuracy and generalization ability. In this paper, we propose an information propagation prediction model, MTGAT, based on the Transformer with modulated position-encoded embedding and graph attention networks. First, the social network graph and the information propagation graph are constructed as a dynamic social network propagation heterogeneous graph. Then, GAT is used to learn the structural features of the users in the heterogeneous graph, while acquiring the time-series features of the users during the message propagation process. Finally, the time-series features are fused with the structural features and location information is added, and the final predicted probability of message propagation is output after Transformer decoding layer processing and probability prediction module. The experimental results show that the MTGAT model outperforms the existing optimal model on all three real datasets.