Curating the Collapse: How Sample-Level Auditing Accelerates Model Collapse in Synthesis of Tabular Data

Abstract

Generative artificial intelligence has fundamentally changed our relationship with data, and the proliferation of artificially generated, synthetic, content on the internet has sparked much concern from the community. Because it is difficult to moderate and identify high-fidelity synthetic examples, future deep learning algorithms and AI systems based on data sourced from the internet are at risk of reciting spurious patterns, defects, and unfairness. Moreover, unchecked synthetic data reuse in generative models drives a self-replicating cycle of deterioration, often under the umbrella term model collapse. In this work, we focus on the synthesis of tabular data, which has received broad attention in high-risk application domains such as healthcare and finance, for dataset augmentation and privacy preservation, but remains understudied for model collapse dynamics. We demonstrate how the effect persists for tabular data and extend our findings to show how various levels of sample curation (output quality filtering) can provoke mode dropping and diversity loss. Finally, we discuss trade-offs associated with auditing and propose future directions for synthetic data generation pipelines.