Iterative Subset Selection for High-fidelity Synthetic Tabular Data

Daniel Gärber, Lea Demelius

Submission Type: Short paper (4 pages)

Keywords: tabular data, synthetic data generation

TL;DR: We present a method for improving synthetic tabular data.

Abstract: We present ISSOSYNTH, an iterative subset selection method for generating high-fidelity synthetic tabular data. The approach won the Mostly AI prize (2025) for generating the synthetic data with the highest fidelity. We evaluate the fidelity, utility and empirical privacy of the approach on various datasets and show that the method is able to improve fidelity and utility on downstream tasks without notably increasing vulnerability to membership inference attacks.

Relevance Comments: synthetic data generation for tabular data

Published Venue And Year: AI for Tabular Data workshop at EurIPS 2025

Submission Number: 44