CasText: Fusion of Text Information Flow and Global Perspective for Predicting the Size of Information Dissemination

Accurately predicting the size of information dissemination has important theoretical and practical significance for formulating content distribution strategies, optimising network resource allocation and conducting effective public opinion management. Current research on the cascade growth size of social media information dissemination mostly focuses on network structure and user behaviour analysis. Still, it neglects the crucial role of textual information in driving information dissemination. We propose a deep learning framework called CasText, which integrates multisource features such as text information, global propagation graphs and local propagation structures to more accurately predict the size of information propagation. Using Sentence-BERT to extract deep semantic features of text and combining it with GNN, precise capture of the interaction between text information and cascading structures has been achieved; using DeepWalk to view the entire social network as a complex graphic structure, high-dimensional feature representations of each social media user can be automatically learned. This global perspective helps to reveal broader dissemination patterns and potential influence paths, thereby improving the accuracy of predicting the size of future information dissemination. In multiple comparative experiments based on a real Weibo cascaded text retweeting dataset, the CasText model improved the MSLE index by 3.1% compared to the baseline model, significantly demonstrating the effectiveness of multisource feature fusion in predicting information dissemination size. We further confirmed the importance of text information, global propagation graphs and local propagation embeddings in improving model performance through ablation experiments.