Prediction of viral spread of digital content based on dynamic engagement graph and graph neural networks

Against the backdrop of the explosive proliferation of information on social media, traditional statistical and dynamic models are constrained by their static assumptions, while deep learning methods rely on artificial features, making it difficult to effectively depict the dynamic evolution of user-content interactions. To address this, the study proposes a framework for predicting the viral spread of digital content based on dynamic engagement graphs and spatio-temporal graph neural networks. This method integrates neighbor features of users and content through graph convolutional structures in the spatial dimension, captures the outbreak, diffusion, and decline processes of spread using gated recurrent mechanisms in the temporal dimension, and highlights the roles of recent and key interactions through time-decay weighting and semantic fusion mechanisms. Experiments show that the proposed method achieves a stable root mean square error of around 0.20 after 200 iterations, representing a reduction of 16–28% compared to other methods. In cross-platform experiments, the proposed method attains a top-twenty hit rate of 0.86, an improvement of 11–19% over other methods, with a rank correlation coefficient increase of over 18%. The results indicate that the proposed method achieves higher prediction accuracy, faster convergence speed, and stronger cross-platform adaptability. The applicability of this method across various datasets and complex scenarios provides new insights and approaches for early identification of viral digital content and public opinion monitoring.