This paper proposes a novel generative diffusion framework, the Structure-prior Informed Diffusion model for Source Localization (SIDSL), for source localization in graph information propagation. While existing deep learning-based approaches struggle with a lack of real-world data, SIDSL leverages topology-based prior information to enable robust source localization even with limited data. Specifically, we apply techniques such as graph label propagation, a GNN-based label propagation module, and diffusion initialization utilizing structured prior information to address unknown propagation patterns, complex topology-propagation relationships, and class imbalance. We effectively transfer knowledge to real-world scenarios by learning pattern-invariant features using synthetic data generated by existing propagation models. Experimental results using four real-world datasets demonstrate that our proposed approach achieves F1 scores that are 7.5-13.3% higher than existing methods, with performance improvements of over 19% and 40%, respectively, in small-data and data-free scenarios.
