Back to search

University of Minnesota

Information-Geometric Generative Models for Mixed Continuous–Categorical Tabular Data

Abstract

dc:description.abstract

Mixed continuous-categorical data pose a geometric problem for continuous generative models. Flow Matching and Gaussian diffusion operate in Euclidean spaces, whereas categorical laws live on probability simplices and can be highly imbalanced. We propose a logit-coordinate framework that embeds categorical variables in smoothed natural-parameter coordinates, concatenates them with transformed numerical variables, and applies continuous generative dynamics in the resulting Euclidean representation. This construction gives a common representation-level formulation for Logit Flow Matching (Logit FM) and Logit Diffusion. To evaluate mixed-data fidelity, we introduce a mixed discrepancy 𝐷ₘᵢₓ that separates categorical marginal error from conditional continuous Wasserstein error. The analysis gives stability bounds showing how vector-field or probability-flow drift error propagates through softmax decoding to mixed-distribution error, and combines these bounds with nonparametric estimation rates that make the effective sample-size loss for rare categories explicit. Controlled simulations show that scaled-logit coordinates improve or match one-hot coordinates within both Flow Matching and Gaussian diffusion, with the clearest gains appearing under severe rare-cell imbalance. Experiments on four mixed-type tabular benchmarks show that Logit FM improves over a one-hot Flow Matching ablation, that a block-conditional Logit FM extension further improves conditional mixed-distribution fidelity, and that Logit Diffusion generally improves over or matches One-Hot Diffusion within the diffusion family. Overall, the results support a representation-level message: respecting the information geometry of categorical variables can improve mixed tabular generation across continuous generative modeling frameworks, while realistic data may also require conditional structure to preserve high-order dependence.

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Shen, Yuefei

Subjects

dc:subject × 5

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Citation
Shen, Yuefei. (2026). Information-Geometric Generative Models for Mixed Continuous–Categorical Tabular Data. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/282402.

Chain of custody

source
Harvested from
University of Minnesota
Base URL
conservancy.umn.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Shen, Yuefei. Information-Geometric Generative Models for Mixed Continuous–Categorical Tabular Data. 2026. https://hdl.handle.net/11299/282402