Abstract
Predicting the timing of information diffusion in social networks is critical for applications ranging from viral marketing to the mitigation of misinformation. This paper introduces R-ODE (Ricci-curvature regulated ODE), a novel framework that treats social networks as dynamic systems. By combining the geometric concept of Ricci curvature with Neural Ordinary Differential Equations (ODEs), R-ODE predicts not just who will be informed by a piece of information, but exactly when. The model treats information flow as a continuous-time trajectory, significantly outperforming existing discrete-time models.
Introduction
Information diffusion prediction—modeling how a tweet, a rumor, or an advertisement spreads—is a cornerstone of social network analysis. Traditionally, models focus on the "who": identifying which user is most likely to be infected next. However, the timing of infection ("when") is just as critical. Knowing the precise timing allows for more effective interventions in spreading processes. Existing "time-aware" models are often limited to predicting the immediate next step. R-ODE aims for personalized time prediction by treating the diffusion process as a continuous-time trajectory in a geometric space.
Methodology
The R-ODE framework conceptualizes the social network as a physical system where information flows along paths of least "transportation effort."
The framework consists of three main components:
- Recursive Graph Convolution: The model builds a temporal graph where users and messages are nodes. It uses an attention mechanism to update user and message embeddings every time a new infection occurs, capturing the evolving state of the system.
- Diffusion via Ricci Curvature: In Riemannian geometry, Ricci curvature quantifies the "effort" required to transport mass between points. R-ODE posits that information naturally flows toward users with the highest Ricci curvature (least transportation effort). This geometric measure replaces traditional, flatter metrics like cosine similarity.
- Neural ODEs for Continuous Dynamics: Instead of treating time as discrete steps, R-ODE defines a "velocity vector field" parameterized by a Graph Neural Network (GNN). By solving this ODE, the model can pinpoint the exact moment a user’s trajectory intersects with a message's influence in the embedding space.
Experimental Results
The authors evaluated R-ODE on three real-world benchmark datasets: Twitter (microblogging), Android (Q&A), and Christianity (Stack Exchange). These datasets provide "cascades"—sequences of users and timestamps of interaction.
As shown in the performance comparison, R-ODE consistently outperforms state-of-the-art baselines like FOREST and DyHGCN. On the Twitter dataset, R-ODE’s Hit@10 score was over 5x higher than earlier models like Deepdiffuse. In time prediction tasks, R-ODE achieved the lowest Root Mean Square Error (RMSE) and proved capable of predicting infection times for users multiple steps ahead in the cascade. Ablation studies confirmed that both the Ricci curvature component and the GNN-based velocity field are essential for these performance gains.
Discussion and Future Work
R-ODE successfully bridges the gap between discrete infection events and the continuous growth of infection potential. By using discrete graph snapshots to "anchor" a continuous ODE trajectory, it provides a more nuanced view of network dynamics.
However, challenges remain. Solving ODEs and calculating Ricci curvature can be computationally expensive for networks with millions of nodes. Additionally, the model's reliance on user features and adjacency matrices means its performance may be sensitive to data sparsity or privacy constraints. Future research could explore alternative types of curvature or apply this framework to other flow-based systems, such as financial transactions.
Summary updated