PDSL: Propagation Dynamics Aware Framework for Source Localization
Researchers combine Graph Neural ODEs with diffusion models to reduce source uncertainty...
Source localization—finding the origin of a viral post, rumor, or infection—is a critical inverse inference problem in network science. Traditional methods struggle because observed outcomes map ambiguously to potential sources, a problem compounded by the stochastic nature of propagation. Most existing approaches focus only on topological uncertainty, ignoring the randomness of the diffusion process itself. The new PDSL framework addresses this by coupling a deep generative model with explicit propagation dynamics, allowing it to approximate the true source distribution while mitigating uncertainty from diffusion stochasticity.
PDSL uses Graph Neural Ordinary Differential Equations to model the continuous-time evolution of information spread without relying on a predefined diffusion mechanism. This flexibility makes it applicable to various propagation scenarios, from social media cascades to epidemic spread. A novel matching mechanism extracts relevant data blocks from the observed network, improving the reliability of source generation. Evaluated on both synthetic benchmarks and real-world diffusion datasets, PDSL consistently outperforms state-of-the-art methods, demonstrating robust performance across diverse applications such as rumor tracing, viral marketing analysis, and disease outbreak detection.
- Integrates deep generative models with propagation dynamics to reduce uncertainty from stochastic diffusion
- Employs Graph Neural ODEs to model continuous propagation without assuming a fixed diffusion mechanism
- Outperforms existing methods on synthetic and real-world datasets, with applications in rumor detection and epidemic tracking
Why It Matters
Better source localization helps professionals trace misinformation, monitor epidemics, and optimize viral marketing strategies.