跳转至

文章背景与核心概要

生成混合类型的表格数据需要同时对多样化的特征分布及其复杂的跨列依赖关系进行建模。传统的变分流匹配虽然能够通过因式分解分布处理不同的端点，但通常将特定特征的处理和跨列交互隐含在共享的主干网络中。

本文介绍了 FUSE（Feature-wise Unified Specialization with cross-column Exchange，即带有跨列交换的特征级统一专业化方法），这是一种旨在明确分离这些角色的创新方法。FUSE 将独立的自适应混合模块分别应用于数值特征和类别特征——允许每个特征组合共享的专业化子网络——同时保持联合注意力机制以维持所有列之间的信息交换。此外，作者还推导了受限条件上下文带来的过剩总体风险，并建立了将连续 Wasserstein 生成误差与端点预测风险相关联的理论界限。在八个表格数据集上的广泛实验表明，FUSE 在分布保真度和下游效用指标上均取得了强大且一致的卓越性能。

FUSE: Feature-Wise Unified Specialization with Cross-Column Exchange for Mixed-Type Tabular Flow Matching

Summary

Generating mixed-type tabular data requires simultaneously modeling diverse feature distributions and their complex cross-column dependencies. While variational flow matching handles distinct endpoints via factorized distributions, it typically leaves feature-specific processing and cross-column interactions implicit within a shared backbone. This paper introduces FUSE (Feature-wise Unified Specialization with cross-column Exchange), a novel approach designed to explicitly separate these roles. FUSE applies separate adaptive mixture modules to numerical and categorical features—allowing each feature to combine shared specialized subnetworks—while maintaining joint attention to preserve information exchange across all columns.

Generating mixed-type tabular data requires simultaneously modeling diverse feature distributions and their complex cross-column dependencies. While variational flow matching handles distinct endpoints via factorized distributions, it typically leaves feature-specific processing and cross-column interactions implicit within a shared backbone. This paper introduces FUSE (Feature-wise Unified Specialization with cross-column Exchange), a novel approach designed to explicitly separate these roles. FUSE applies separate adaptive mixture modules to numerical and categorical features—allowing each feature to combine shared specialized subnetworks—while maintaining joint attention to preserve information exchange across all columns.

Paper Metadata

  • arXiv ID: arXiv:2608.07294 [cs.LG]
  • Subject Area: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
  • Authors: Suman Cha, Seongchan Lee, Dohyun Ko, Hyunjoong Kim
  • Submitted: August 7, 2026
  • Length: 19 pages, 7 figures, 7 tables
  • License: Creative Commons Attribution 4.0 International (CC BY 4.0) <img alt="license icon" role="presentation" src=".././images/345c7ad61f1b.png" />
  • arXiv ID: arXiv:2608.07294 [cs.LG]
  • Subject Area: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
  • Authors: Suman Cha, Seongchan Lee, Dohyun Ko, Hyunjoong Kim
  • Submitted: August 7, 2026
  • Length: 19 pages, 7 figures, 7 tables
  • License: Creative Commons Attribution 4.0 International (CC BY 4.0) <img alt="license icon" role="presentation" src=".././images/345c7ad61f1b.png" />

Abstract

生成混合类型的表格数据需要联合建模多样化的特征分布及其复杂的跨列依赖关系。变分流匹配通过因式分解分布处理不同的端点，但仍将特定特征的处理和跨列交互隐含在共享的主干网络中。我们引入了带有跨列交换的特征级统一专业化方法（FUSE），以明确分离这些角色。FUSE 将独立的自适应混合模块应用于数值特征和类别特征，使每个特征能够组合共享的专业化子网络，同时利用联合注意力机制保持所有列间的信息交换。我们还刻画了受限条件上下文所带来的过剩总体风险，并建立了将连续 Wasserstein 生成误差与端点预测风险相联系的界限。在八个表格数据集上的全面实验表明，FUSE 在分布保真度和下游效用指标上均实现了强大且一致的性能表现。

Generating mixed-type tabular data requires jointly modeling diverse feature distributions and their complex cross-column dependencies. Variational flow matching handles distinct endpoints via factorized distributions, yet leaves feature-specific processing and cross-column interactions implicit within a shared backbone. We introduce Feature-wise Unified Specialization with cross-column Exchange (FUSE) to explicitly separate these roles. FUSE applies separate adaptive mixture modules to numerical and categorical features, allowing each feature to combine shared specialized subnetworks, while joint attention preserves information exchange across all columns. We also characterize the excess population risk from restricted conditioning contexts and bound the continuous Wasserstein generation error by endpoint-prediction risk. Comprehensive experiments on eight tabular datasets demonstrate that FUSE achieves strong and consistent performance across distributional fidelity and downstream utility metrics.

Key Contributions & Architecture Highlights

  • 显式角色分离： 将特定特征的处理与跨列交互解耦，摆脱了纯粹隐式的共享主干网络架构。
  • 自适应混合模块： 部署专门针对数值和类别特征定制的自适应混合模块，使它们能够动态组合共享的专业化子网络。
  • 联合注意力机制： 保持并促进了所有表格列之间的强大信息交换。
  • 理论保证： 刻画了由受限条件上下文引起的过剩总体风险，并确立了将连续 Wasserstein 生成误差与端点预测风险联系起来的界限。
  • 实证验证： 通过对八个表格数据集的全面实验得到验证，在分布保真度和下游效用指标上均表现出持续的优越性。
  • Explicit Role Separation: Disentangles feature-specific processing from cross-column interactions, moving away from purely implicit shared-backbone architectures.
  • Adaptive Mixture Modules: Deploys dedicated adaptive mixture modules tailored to numerical and categorical features, enabling them to dynamically combine shared specialized subnetworks.
  • Joint Attention Mechanism: Preserves and facilitates robust information exchange across all table columns.
  • Theoretical Guarantees: Characterizes the excess population risk stemming from restricted conditioning contexts and establishes bounds linking continuous Wasserstein generation error to endpoint-prediction risk.
  • Empirical Validation: Demonstrated through comprehensive experiments across eight tabular datasets, showing consistent superiority in both distributional fidelity and downstream utility metrics.

Full-Text & Access Links