MGOIDP: Memory-Enhanced Graph Neural Ordinary Differential Equations for Microscopic Information Diffusion Prediction

The advent of social networking has transformed how information spreads, making it crucial to accurately model and predict microscopic information diffusion. Recent research has focused on refining dynamic user representations by examining relationships and roles within both social and diffusion graphs. However, existing dynamic graph-based approaches typically rely on discrete sequences of snapshots rather than modeling the continuous evolution of users' hidden states, a limitation that often yields suboptimal predictive accuracy. To address this limitation, this paper proposes MGOIDP, Memory-enhanced Graph neural Ordinary differential equations for microscopic Information Diffusion Prediction, which models the continuous-time evolution of information diffusion on both social graphs and diffusion graphs. Specifically, MGOIDP employs two distinct graph ODEs to effectively capture the different properties of these graphs. For the diffusion graph, which contains fewer edges and better represents potential propagation pathways, a dynamic edge-coupled graph ODE is introduced to refine node interactions over time. For the social graph, which has a high density of edges and imposes high computational costs, a scalable diffusion-based graph ODE is incorporated to efficiently model the information diffusion process. Furthermore, memory-enhanced representations are generated by retrieving context-aware user embeddings from temporally adjacent ODE states, ensuring that the current preferences of users are effectively encoded to improve prediction accuracy. Finally, MGOIDP utilizes a multi-head attention module to predict the next user in the diffusion cascade. Experiments on four real-world datasets show its superior performance over state-of-the-art models in Hits@K and MAP@K metrics.