Researchers have developed a new method for generating discrete graphs that addresses limitations in existing diffusion models. These models typically corrupt graph structures through independent edge updates, leading to an Erdős--Rényi (E-R) random graph as the terminal prior. This fundamental issue hinders the ability of graph neural networks (GNNs) to effectively denoise and recover meaningful node heterogeneity and higher-order structure, especially without node features. The proposed solution introduces a Latent Sociability Prior (LSP) and a joint diffusion process that simultaneously evolves latent node sociability variables and graph structure. This approach removes edge independence while maintaining exchangeability, guiding edge updates and preserving node heterogeneity without complex encoders. Experiments on various graph generation datasets demonstrate that this framework more accurately captures graph distributions compared to baseline methods that do not utilize node features. AI
IMPACT Introduces a novel approach to graph generation that could improve the accuracy and expressiveness of models used in various AI applications.
RANK_REASON Academic paper detailing a new method for graph generation. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
