<p>Popularity prediction, which aims to predict the diffusion size of a information cascade, plays a crucial role in understanding the diffusion dynamics and enabling various applications in social networks. Existing methods primarily rely either on the generative point process such as Poisson and Hawkes processes, or deep graph modeling on the static diffusion graph. However, they overlook the fine-grained diffusion dynamics on both node-level and graph level, and hardly generalize to new users and scenarios. Moreover, they treat each cascade as an individual case, neglecting the correlations with other similar cascades. To address these, in this paper we propose a novel feature-driven dynamic graph learning model named FedDGL for accurate cascade popularity prediction. FedDGL first divides the observation time into finite time slots, and extracts multiple microscopic and macroscopic diffusion features that describe the diffusion dynamics of each node and cascade graph respectively. Then, a microscopic features-driven cascade graph learning strategy is proposed to learn the node-level cascade graph embedding, which is further integrated with the graph-level macroscopic features for comprehensive cascade graph embedding. Subsequently, a Bidirectional Long Short-Term Memory (BiLSTM) model is utilized to to learn the temporal evolutionary diffusion dynamics representation and predict the cascade-based popularity. Additionally, we also propose a well-designed popularity-supervised Auto-encoder model to embed all cascade graphs and retrieve most similar cascades with associated popularity as referred prediction. Finally, the two predicted popularity are adaptively fused to obtain the ultimate diffusion size. Extensive experiments on real-world datasets demonstrate the effectiveness of our model compared to state-of-the-art methods, while enabling good generalization ability.</p>