Tabmosaic generates realistic missing data for any table without retraining
TabMosaic: Unified In-Context Diffusion for Tabular Generative Tasks
Machine LearningArtificial Intelligence
Summary
Many methods can fill in missing pieces or create new data for tables, but they often need to be retrained for each table or only handle one missing piece at a time. The authors present TabMosaic, a new approach that treats these tasks as one problem: predicting any group of missing cells from the rest of the table. Their technique uses a method called diffusion, which predicts multiple missing values together, not one by one. This allows TabMosaic to work on unseen tables without extra training and perform better across many real-world datasets.
What this means in practice
- •For data engineering teams: Automatically fill in missing or partial data in diverse tabular sources without needing to retrain new models per dataset.
- •For business intelligence developers: Generate realistic synthetic tabular records conditioned on specified values to enhance data augmentation and analysis workflows.
Authors
Jayoung Kim, Sanghyun Hong, Noseong Park
Abstract
Generative modeling for tabular data has advanced to the point where models can synthesize realistic records, recover missing values, and generate samples conditioned on class labels or observed attributes. Despite recent advances, existing approaches either require table-specific training or restrict in-context learning to single-cell prediction. In this work, we formulate these tabular generative tasks as a single problem: modeling the joint conditional distribution of an arbitrary set of target cells given the observed cells of an unseen table. To this end, we turn to diffusion, whose denoising objective is defined jointly over all target cells rather than one cell at a time. Under the pretraining distribution, this objective is minimized by the posterior predictive distribution over the target set. Building on this, we propose TabMosaic, a unified in-context diffusion model for generating mixed-type target cells in unseen, potentially incomplete tables without table-specific training. We further show that its pretraining objective bounds the conditional negative log-likelihood and the divergence from the posterior predictive for arbitrary target sets, extending the Bayesian account of in-context learning from a single-value prediction to joint multi-cell generation. Across 48 real-world tables, TabMosaic ranks first on all evaluated generative tasks against 20 baseline approaches