Mining user features with hyperbolic representations for diffusion prediction.

The study of information diffusion has great practical significance in real-world applications, and has attracted the attention of many scholars. Despite remarkable progress, existing diffusion prediction models still suffer from several key limitations: (i) Euclidean-based embeddings fail to capture the inherent hierarchical structures of social networks; (ii) most models overlook the layer-by-layer propagation dynamics and the mutual influence between adjacent users, leading to suboptimal behavioral modeling. To address these shortcomings, we propose a novel deep learning framework, named Hyper-MUF, which leverages hyperbolic representations to mine both static and dynamic user features for diffusion prediction. Specifically, the model first reconstructs the social network in a hyperbolic space to capture hierarchical user relationships. In addition to the static user representations in social networks, we also use pooling and attention mechanisms to mine user dynamic features in cascade sequences to fit the diffusion process. By integrating these two aspects, Hyper-MUF provides a unified representation that reflects both structural hierarchy and temporal evolution in information diffusion. We validated the performance of our model on four real datasets and demonstrated the effectiveness of each module in this method. In large-scale datasets, the prediction performance of our model is improved by 10% to 20% compared with the state-of-the-art methods, showing the superiority of the model.