A small subset of training samples drives most memorization in tabular diffusion models, and pruning them based on early memorization signals reduces measured leakage, though the evaluation metric makes part of the gain mechanical.
Diffusion probabilistic models generalize when they fail to memorize
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
A Closer Look on Memorization in Tabular Diffusion Model: A Data-Centric Perspective
A small subset of training samples drives most memorization in tabular diffusion models, and pruning them based on early memorization signals reduces measured leakage, though the evaluation metric makes part of the gain mechanical.