Active Tabular Augmentation via Policy-Guided Diffusion Inpainting

2026-05-11

AI Summary

Existing tabular data augmentation methods often prioritize distributional fidelity at the expense of downstream model performance. To address this limitation, this work proposes TAP, a novel framework that shifts the augmentation objective from fidelity to utility-driven generation. TAP employs diffusion-based inpainting to synthesize samples and introduces a lightweight policy network that dynamically guides the generation process toward producing the most valuable data based on the current learner’s state. Additionally, it incorporates an explicit gating mechanism coupled with a sliding-window commitment strategy to ensure safe and conservative sample injection. Extensive experiments across seven real-world datasets demonstrate that TAP consistently outperforms existing approaches, achieving up to a 15.6% improvement in classification accuracy and a 32% reduction in RMSE for regression tasks.

Abstract

Generative tabular augmentation is appealing in data-scarce domains, yet the prevailing focus on distributional fidelity does not reliably translate into better downstream models. We formalize a fidelity-utility gap: common generative objectives prioritize distributional plausibility, whereas augmentation succeeds only when injected samples reduce the current learner's held-out evaluation loss. This gap motivates learning not just how to generate, but what to generate and when to inject as training evolves. We propose TAP (Tabular Augmentation Policy), which couples diffusion inpainting with a lightweight, learner-conditioned policy to steer generation toward high-utility regions and controls safe injection via explicit gating and conservative windowed commitment. Under severe data scarcity, TAP consistently outperforms strong generative baselines on seven real-world datasets, improving classification accuracy by up to 15.6 percentage points and reducing regression RMSE by up to 32%.

Problem

• tabular augmentation
• data scarcity
• fidelity-utility gap
• generative modeling
• downstream performance

Innovation

• policy-guided generation
• diffusion inpainting
• tabular data augmentation
• utility-aware generation
• learner-conditioned policy