Home | Publications | ATG+26

Tabular-DC-Bench: A Large-Scale Meta-Dataset for Hyperparameter Optimization in Deep Clustering

MCML Authors

Link to Profile Thomas Seidl PI Matchmaking

Thomas Seidl

Prof. Dr.

Core PI

Database Systems, Data Mining and AI

Abstract

While Automated Machine Learning (AutoML) has made progress in supervised domains, deep clustering for tabular data has not gained comparable traction. This is largely due to the difficulty of hyperparameter optimization and model selection in the absence of ground-truth validation signals. We aim to address this gap by introducing Tabular-DC-Bench, a large-scale meta-dataset for deep clustering research on tabular data. We curated 100 tabular datasets across varying scales and evaluated five representative deep clustering algorithms using 75,000 configurations, executing each run five times to quantify stability. The resulting dataset contains 375,000 individual model evaluations, including ground-truth performance, internal metrics, and training loss curves. We demonstrate the utility of the dataset by analyzing the failure of common unsupervised validation proxies and by quantifying the performance gap between default and optimized configurations across different data regimes.

inproceedings ATG+26

ABCD Track @AutoML 202

Track on Applications, Benchmarks, Challenges, and Datasets at the International Conference on Automated Machine Learning. Ljubljana, Slovenia, Sep 28-Oct 01, 2026. To be published. Preprint available.

Authors

M. Aljoud • G. M. Tavares • S. Gilhuber • C. Leiber • T. Seidl

Links

URL

Research Area

 A3 | Computational Models

BibTeXKey: ATG+26

#p-seidl
Back to Top