Researchers have analyzed continuous normalized flow models parameterized by graph neural networks (GNNs) to understand how structural errors affect graph signal generation. They derived explicit stability bounds, demonstrating how perturbations in graph structure influence the final sampled signals. To enhance robustness, they introduced a stability-promoting regularized flow matching strategy that penalizes the spatial Lipschitz constant during training, showing improved performance on synthetic and real-world fMRI data. AI
RANK_REASON Academic paper on generative models for graph signals. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →