TabularStudy

Code for Tabular Learning Revisited: An Empirical Study of Tabular Classification.

The paper has been accepted for publication in Transactions on Machine Learning Research (TMLR). The OpenReview page is available at: https://openreview.net/forum?id=I8BIGp4XOb

Overview

This repository contains the code used to run the experiments in our paper.

Note: This repository is not intended to be a general‑purpose library.
It is a collection of scripts that reproduce the experimental pipeline.

Experiments are conducted on the OpenMLCC18 benchmark suite.
Four datasets were excluded due to persistent memory issues.
The final list of dataset IDs is provided in openmlcc18_tasks.txt.

Each method resides in its own folder that contains everything required to run it on the OpenMLCC18 benchmark.

This repository includes code adapted from third-party projects. Their original licenses are retained in the corresponding subdirectories.

---

Setup

Create the Conda environment defined in environment.yml:

conda env create -f environment.yml
conda activate tabularstudy

Most experiment scripts log to Weights & Biases. To run them without a W&B account or network logging, set offline mode before launching a script:

export WANDB_MODE=offline

This environment was used for the following methods:

• CatBoost
• XGBoost
• LightGBM
• FT-Transformer
• MLP-PLR
• ResNet
• SAINT
• TabNet
• XTab

The preprocessing and training pipeline are largely based on the original FT-Transformer codebase.

Other methods (e.g. AutoML libraries or foundation models) have their own dependencies and preprocessing.
Please follow the official instructions of each method to create the correct environment.

ModernNCA uses the external TALENT package and is not vendored in this repository. Install TALENT separately before running ModernNCA/run_modernnca.py.

pip install TALENT

---

Quick-start (single fold, single dataset)

To sanity-check your installation, you can run ResNet on the Balance-Scale dataset
(OpenML ID = 11) for outer fold 0 with 100 Optuna trials:

cd ResNet
python run_resnet.py \
--seed 0 \
--outer_fold 0 \
--dataset 11 \
--n_trials 100 \
--experiment_name "ResNet_HPO" \
--tune

Omit--tune to run with default hyper-parameters.

---

Entry-Point Scripts

Method	Script Path
AutoGluon	autogluon/run_autogluon.py
CARTE	carte/examples/carte_single_tables.py
CatBoost	Catboost/run_catboost.py
FT-Transformer	FT_Transformer/bin/run_ft.py
MLP-PLR	MLP/run_mlp.py
LightGBM	LightGBM/run_lightgbm.py
ModernNCA	ModernNCA/run_modernnca.py
RealMLP	RealMLP/run_realmlp.py
TabM	RealMLP/run_tabm.py
ResNet	ResNet/run_resnet.py
SAINT	saint/run_saint.py
TabNet	TabNet/run_tabnet.py
TabICL	TabICL/run_tabicl.py
TabPFN	TabPFN/run_tabpfn.py
TabPFNv2	TabPFNv2/run_tabpfnv2.py
TPBERTa	TPBerta/scripts/finetune/default/run_default_config_tpberta.py
XGBoost	XGBoost/run_xgboost.py
XTab	XTab/run_xtab_finetune.py

---

Parallelisation & Nested Cross-Validation

Experiments use nested cross-validation.
The outer loop is parallelised via Slurm job arrays.

Example Slurm script for TabM

#!/bin/bash#SBATCH -t 1-00:00 # Runtime (D-HH:MM)#SBATCH --gres=gpu:1 # Number of GPUs#SBATCH -o log/%x.%N.%j.out # STDOUT#SBATCH -e log/%x.%N.%j.err # STDERR#SBATCH -J JOB_NAME # Job name#SBATCH -a 0-679 # 68 datasets × 10 outer folds# Load environmentsource~/.bashrc
conda activate ENVIRONMENT_NAME
# (Optional) Proxy for outbound traffic# export https_proxy=http://proxy:portecho"Working dir : $PWD"echo"Started at : $(date)"echo"Job : $SLURM_JOB_NAME"echo"CPUs/node : $SLURM_JOB_CPUS_PER_NODE"echo"Job ID : $SLURM_JOB_ID"echo"Partition : $SLURM_JOB_PARTITION"# Read dataset listread -a DATASETS < openmlcc18_tasks.txt
# Calculate dataset index & fold
DATASET_INDEX=$((SLURM_ARRAY_TASK_ID /10))
OUTER_FOLD=$((SLURM_ARRAY_TASK_ID %10))
DATASET=${DATASETS[$DATASET_INDEX]}# Run TabM
python run_tabm.py \
--seed 0 \
--outer_fold $OUTER_FOLD \
--dataset $DATASET \
--n_trials 100 \
--experiment_name "TabM_HPO" \
--tune
echo"DONE"echo"Finished at : $(date)"

---

Citation

If you find this work useful, please cite:

@article{zab{\"e}rgja2026tabular,
title={Tabular Learning Revisited: An Empirical Study of Tabular Classification},
author={Guri Zab{\"e}rgja and Arlind Kadra and Christian Frey and Josif Grabocka},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2026},
url={https://openreview.net/forum?id=I8BIGp4XOb},
note={}
}