跳转至

文章背景与核心概要

合成表格数据生成器在隐私保护、数据增强等领域扮演着重要角色。传统上,评估这些模型不仅需要考察它们保持单列边缘分布的能力,还需要关注它们捕捉复杂列间依赖关系(inter-column dependencies)的表现。然而,现有的标准认证指标(如基于逻辑回归的分类器双样本检验 C2ST)在检测列间关系缺陷时往往力不从心。

为了解决这一痛点,该研究引入了一套先进的诊断框架,利用梯度提升树(gradient-boosted)C2ST,将评估得分巧妙地分解为边缘项、依赖项以及数值-类别交叉项。当该方法被应用于现代流匹配(TabbyFlow)和扩散(TabDiff)模型时,研究人员揭示了一个普遍存在的“依赖鸿沟”(dependency gap)。该研究进一步证明,这一鸿沟并非由采样离散化、模型容量限制或平均场目标函数引起,而是源于训练过程中对直接依赖监督的根本性缺失。


衡量依赖鸿沟:诊断表格生成模型中的列间保真度 (Measuring the Dependency Gap: Diagnosing Inter-Column Fidelity in Tabular Generative Models)

作者: Jie Zhang
主要学科: 机器学习 (cs.LG),兼跨人工智能 (cs.AI)
arXiv ID: 2607.21636v4 [cs.LG]
DOI: 10.48550/arXiv.2607.21636
提交历史: 2026年7月20日提交;2026年8月19日最后修订。


📌 摘要与总结 (Summary)

Synthetic tabular data generators are traditionally evaluated on their ability to preserve both column-wise marginal distributions and complex inter-column dependencies. However, standard certification metrics—such as the logistic-regression-based classifier two-sample test (C2ST)—are largely ineffective at detecting flaws in inter-column relationships.

To address this, the author introduces an advanced diagnostic framework utilizing a gradient-boosted C2ST, decomposing evaluation scores into marginal, dependency, and cross-terms. When applied to modern flow-matching (TabbyFlow) and diffusion (TabDiff) models, the approach uncovers a persistent "dependency gap." The study proves this gap is not merely an artifact of sampling discretization, model capacity limits, or mean-field objectives, but rather stems from a fundamental lack of direct dependency supervision during training.

传统的合成表格数据生成器通常通过其保持单列边缘分布(column-wise marginal distributions)以及复杂的列间依赖关系的能力来评估。然而,标准的认证指标——例如基于逻辑回归的分类器双样本检验(C2ST)——在检测列间关系的缺陷时基本上形同虚设。

为了解决这一问题,作者引入了一个先进的诊断框架,该框架利用梯度提升 C2ST,将评估得分分解为边缘项、依赖项和交叉项。当应用于现代流匹配(TabbyFlow)和扩散(TabDiff)模型时,该方法揭示了一个持久存在的“依赖鸿沟”。研究证明,这一鸿沟既不是采样离散化、模型容量限制或平均场目标的产物,而是源于训练过程中根本上缺乏直接的依赖监督。


📄 论文摘要 (Abstract)

Synthetic tabular data are valued for preserving not just column-wise marginals but inter-column dependency. Yet the most commonly reported certification score, a linear (logistic-regression) classifier two-sample test (C2ST), is largely blind to it: a fully-factorized baseline that destroys all inter-column dependency still appears nearly real, a known weakness we confirm on four benchmarks, while pairwise Trend penalizes the same baseline only mildly.

We therefore apply a stronger, gradient-boosted C2ST and decompose its score into marginal, dependency, and numerical-categorical cross terms, each read against a zero-dependency reference and a real-data oracle. Applied to flow-matching (TabbyFlow) and diffusion (TabDiff) generators, it exposes a persistent dependency gap of the same order in both. Destroying dependency outright with every marginal intact collapses minority-class F1 by 0.38-0.61, though the generators' much smaller residual gaps do not track the shortfalls that remain. The gap is neither a structural limitation of mean-field objectives nor an artifact of sampling discretization, and a 16x capacity increase does not close it. Shrinking capacity eightfold, however, doubles it, so the measurement does respond to capacity; what remains points to the absence of direct dependency supervision.

合成表格数据之所以受到重视,不仅是因为它们能保留单列边缘分布,更是因为它们能保留列间依赖关系。然而,最常报告的认证得分——线性(逻辑回归)分类器双样本检验(C2ST)——对这一点基本上是盲目的:一个破坏了所有列间依赖关系的完全因式分解基线看起来仍然几乎是真实的(我们在四个基准上证实了这一已知弱点),而成对趋势(pairwise Trend)对同一基线的惩罚也很轻微。

因此,我们应用了一种更强大的、基于梯度提升的 C2ST,并将其得分分解为边缘项、依赖项以及数值-类别交叉项,每一项都对照零依赖参考基准和真实数据神谕(oracle)进行解读。将其应用于流匹配(TabbyFlow)和扩散(TabDiff)生成器时,它在两者中都暴露出了相同数量级的持久依赖鸿沟。彻底破坏依赖关系但保持所有边缘分布不变,会使少数类 F1 分数崩塌 0.38-0.61,尽管生成器较小的残余鸿沟并没有追踪到剩余的性能短板。该鸿沟既不是平均场目标的结构性限制,也不是采样离散化的产物,容量增加 16 倍也无法消除它。然而,将容量缩小八倍会使鸿沟翻倍,因此该测量指标确实对容量敏感;所剩下的问题指向了训练中直接依赖监督的缺失。