Your Discrete Graph Diffusion Model is Secretly Memorizing
The paper says memorization appears only when the model is asked to denoise a corrupted training graph.
Jamadandi and Keriven test discrete graph diffusion models trained on two-block stochastic block models and find a strong train-data bias in reverse diffusion. Sample-level metrics during normal generation do not flag the issue. The model reconstructs corrupted training graphs far better than held-out graphs, while the generalization gap grows. Early stopping does not fix it because memorization starts before generated graphs show community structure. ArXiv · AI/CL/LG's note
Jamadandi and Keriven test discrete graph diffusion models trained on two-block stochastic block models and find a strong train-data bias in reverse diffusion. Sample-level metrics during normal generation do not flag the issue. The model reconstructs corrupted training graphs far better than held-out graphs, while the generalization gap grows. Early stopping does not fix it because memorization starts before generated graphs show community structure. ArXiv · AI/CL/LG's note
score 5