THDiff: Joint modeling of social and temporal factors in hyperbolic space for time-aware information diffusion prediction

Information Diffusion Prediction (IDP) aims to forecast how information spreads among users on social platforms at a micro level. Most existing methods emphasize social relations while overlooking fine-grained temporal dynamics, even though both factors are essential to diffusion. Moreover, both social and temporal factors exhibit hierarchical and imbalanced structures, posing challenges for effective modeling via conventional Euclidean methods. In particular, our analysis of real-world diffusion data reveals highly irregular and heavy-tailed time-interval distributions, suggesting the need for refined representations that can accommodate such imbalanced patterns. While hyperbolic geometry has demonstrated strong capability in capturing hierarchical social influence, its potential for modeling temporal imbalances remains largely unexplored. Motivated by these observations, we propose a novel time-aware hyperbolic diffusion prediction model (THDiff) that jointly models social and temporal factors within a unified hyperbolic framework. Specifically, THDiff uses a rotation-assisted hyperbolic embedding to construct user representations that encode global hierarchical social influence. Building on these representations, it leverages hyperbolic attention to construct cascade representations from both social and temporal perspectives and fuses them through a hyperbolic Hawkes process, enabling comprehensive modeling of complex dynamic diffusion dependencies. By jointly modeling social and temporal factors, THDiff effectively predicts not only who will be activated but also when diffusion events are likely to occur. Extensive experiments across four real-world datasets demonstrate that the proposed model outperforms existing methods, achieving up to 8.2% and 10.9% improvements in Hits@1 and MRR, respectively.