Influence-Aware Dynamic Graph Learning for Popularity Prediction

Abstract

With the rapid spread of information, popularity prediction has become a critical problem in decision-making, trend analysis, and recommendation. Recent studies address this problem by modeling historical interaction sequences as dynamic graphs and learning node representations to characterize popularity evolution. However, these methods typically rely on neighborhood aggregation, which biases learning toward topological locality and may miss long-range dependencies when distant entities exhibit similar evolution patterns. In fact, node popularity is driven by multiple influence factors whose effects can propagate across different graph regions and evolve beyond immediate neighborhoods. To address this limitation, we propose a cross-topology, influence-aware dynamic graph learning framework that explicitly models latent influence factors from a global perspective. Firstly, our approach jointly captures interaction density, temporal growth dynamics, and evolving structural states to form influence-aware node representations. Secondly, node dynamics are routed to latent influence factors whose representations are continuously updated over time to capture shared popularity evolution across different graph regions. Finally, we aggregate influence patterns along node trajectories across multiple time steps, enabling effective and efficient modeling of long-range dependencies. Experiments on four real-world datasets with diverse popularity distributions demonstrate the effectiveness of our model and its clear advantages over locality-aware neighborhood aggregation. Beyond improved prediction accuracy, our approach provides interpretable insights into the underlying factors governing popularity dynamics, offering a new perspective for high-order representation learning in dynamic graphs.