MedFeat: Model-Aware and Explainability-Driven Feature Engineering with LLMs for Tabular Prediction

Zizheng Zhang; Yiming Li; Justin Xu; Jinyu Wang; Rui Wang; Lei Song; Jiang Bian; David W Eyre

cs.LGFeb 10, 2026

MedFeat: Model-Aware and Explainability-Driven Feature Engineering with LLMs for Tabular Prediction

Authors: Zizheng Zhang, Yiming Li, Justin Xu, Jinyu Wang, Rui Wang, Lei Song, Jiang Bian, David W Eyre, +1 more

Organizations: Microsoft Research · University of Oxford

Abstract

In clinical tabular prediction, classical machine learning models with feature engineering often outperform neural methods. LLMs are increasingly used to automate this process, acting as domain experts that propose diverse feature transformations to boost downstream performance. However, the feature generation process of existing LLM-based methods is agnostic to the downstream learner: the LLM receives no signal about which features currently drive predictions or where the model's representational capacity falls short, so proposals are neither targeted to promising regions of the feature space nor tailored to the learner's inductive bias. This shortcoming is amplified in healthcare data, which simultaneously exhibits class imbalance, heterogeneous feature spaces, and strict interpretability requirements. In this paper, we propose MedFeat, the first feature engineering framework inspired by the workflow of machine learning practitioners, leveraging model-awareness and feature importance signals to iteratively guide feature discovery for clinical tabular learning. We evaluate MedFeat on a broad range of challenging real-world clinical tasks and show that it statistically significantly outperforms state-of-the-art baselines, with an average F1 improvement of more than 10% over the baseline across models with distinct inductive biases.

Figures & tables

Figure 1: Overview of MedFeat. MedFeat iteratively augments clinical tabular data with LLM-generated features. (1) Starting from a model trained on raw features, the framework computes feature importance. (2) It then forms importance-weighted “feature islands” by sampling a subset of influential features, (3) and guides the LLM with the island profiles, model constraints, and a memory bank to propose clinically meaningful transformations. (4) Proposed features are validated and evaluated locally, (5) accepted features are added to the data, and the feedback memory is updated for the next iteration.

Method

Mortality-ICU

Mortality-10yrs

Heart Failure

Mortality-Ward

Cog-Impairment

Discharge

Readmission

Avg. Rank

\downarrow

AUROC

XGBoost

Baseline

9.5

\pm

0.9

75.6

\pm

1.3

23.6

\pm

0.8

68.3

\pm

1.9

12.9

\pm

0.6

67.4

\pm

1.2

1.4

\pm

0.2

71.0

\pm

1.4

55.3

\pm

1.7

73.8

\pm

1.6

39.8

\pm

0.4

71.0

\pm

0.5

29.7

\pm

0.5

58.7

\pm

1.7

4.1

OpenFE

10.2

\pm

0.5

73.8

\pm

0.5

25.0

\pm

1.8

69.2

\pm

1.9

12.2

\pm

0.4

65.6

\pm

1.0

6.1

\pm

4.1

74.8

\pm

1.1

56.7

\pm

1.1

74.7

\pm

1.0

38.1

\pm

0.8

68.9

\pm

1.1

29.2

\pm

0.3

59.0

\pm

0.4

3.6

CAAFE

10.0

\pm

0.7

75.8

\pm

1.4

24.9

\pm

1.6

68.2

\pm

2.1

12.8

\pm

0.6

67.9

\pm

0.7

1.5

\pm

0.5

68.3

\pm

3.2

55.6

\pm

1.3

74.1

\pm

1.4

40.2

\pm

0.8

71.1

\pm

0.7

30.0

\pm

0.2

60.4

\pm

0.7

3.2

FeatLLM

N/A

50.0

\pm

0.0

N/A

55.9

\pm

5.5

N/A

50.0

\pm

0.0

N/A

50.0

\pm

0.0

N/A

51.3

\pm

1.9

N/A

50.0

\pm

0.0

N/A

50.0

\pm

0.0

6.0

Table 1: Performance improvement by MedFeat across 7 clinical prediction tasks. Bold indicates best and shaded indicates second-best method. Avg. Rank computed across AUROC, F1, and AUPRC. MedFeat achieves the best performance under both classifiers significantly. Improv. indicates improvement of MedFeat vs. the raw baseline. Mean and standard deviation over 5 random splits.

Model	F1			AUROC			AUPRC
Model	MedFeat	Baseline	$\Delta$ % [min, max]	MedFeat	Baseline	$\Delta$ % [min, max]	MedFeat	Baseline	$\Delta$ % [min, max]
Logistic Reg.	26.6 $\pm$ 1.1	25.8 $\pm$ 1.1	$5.5\ [1.5,\ 20.7]$	73.0 $\pm$ 1.3	72.2 $\pm$ 1.2	$1.1\ [0.1,\ 2.3]$	23.5 $\pm$ 0.9	22.6 $\pm$ 0.9	$4.8\ [1.8,\ 19.0]$
XGBoost	27.8 $\pm$ 1.1	27.3 $\pm$ 1.0	$3.1\ [0.4,\ 8.7]$	75.3 $\pm$ 1.4	74.8 $\pm$ 1.4	$0.6\ [0.2,\ 1.8]$	27.5 $\pm$ 1.8	27.0 $\pm$ 1.7	$3.8\ [0.5,\ 16.2]$
Overall	27.2 $\pm$ 1.1	26.6 $\pm$ 1.0	$\mathbf{4.3\ [0.4,\ 20.7]}$	74.1 $\pm$ 1.4	73.5 $\pm$ 1.3	$\mathbf{0.8\ [0.1,\ 2.3]}$	25.5 $\pm$ 1.4	24.8 $\pm$ 1.3	$\mathbf{4.3\ [0.5,\ 19.0]}$

Table 2: Performance summary across models and metrics(%) after extensive HPO . Metrics are averaged over all 7 tasks and 2 models.

\Delta

% = mean relative improvement; brackets show [min, max] range across tasks. Mean and standard deviation over 5 random splits.

Ablation Mode	F1 ( $\Delta$ %)	AUROC ( $\Delta$ %)	AUPRC ( $\Delta$ %)
w/o SHAP Importance	25.3 $\pm$ 1.0 (–4.5%)	71.3 $\pm$ 1.2 (–1.5%)	22.8 $\pm$ 0.9 (–4.2%)
w/o Model-Awareness	25.6 $\pm$ 1.0 (–3.4%)	71.0 $\pm$ 1.2 (–1.9%)	23.2 $\pm$ 1.0 (–2.5%)
w/o Island Sampling	25.8 $\pm$ 1.1 (–2.6%)	71.1 $\pm$ 1.3 (–1.8%)	23.3 $\pm$ 1.0 (–2.1%)
w/o Memory	25.8 $\pm$ 1.0 (–2.6%)	71.1 $\pm$ 1.2 (–1.8%)	23.3 $\pm$ 1.0 (–2.1%)

Table 3: Ablation study summary. Values in parentheses denote performance drop (%) relative to the full model. Mean and standard deviation over 5 random splits.

	CAAFE			OCTree			FeatLLM		MedFeat (Ours)
LLM Backbone	F1	AUROC	AUPRC	F1	AUROC	AUPRC	AUROC	AUPRC	F1	AUROC	AUPRC
gpt-4o-2024-11-20	25.1 $\pm$ 0.9	70.8 $\pm$ 1.3	22.5 $\pm$ 1.1	25.4 $\pm$ 1.0	71.0 $\pm$ 1.2	22.5 $\pm$ 0.9	52.2 $\pm$ 1.6	12.2 $\pm$ 0.7	26.5 $\pm$ 1.0	72.4 $\pm$ 1.4	23.8 $\pm$ 0.9
gpt-5.4	25.2 $\pm$ 1.0	70.7 $\pm$ 1.2	22.5 $\pm$ 1.0	25.8 $\pm$ 1.0	70.6 $\pm$ 0.9	21.8 $\pm$ 0.8	53.8 $\pm$ 2.8	13.3 $\pm$ 1.4	26.5 $\pm$ 0.9	72.2 $\pm$ 1.2	24.8 $\pm$ 1.3
gemini-3.1-pro-preview	25.2 $\pm$ 0.8	70.8 $\pm$ 1.3	22.5 $\pm$ 0.9	25.8 $\pm$ 1.0	70.8 $\pm$ 1.1	21.9 $\pm$ 0.8	53.5 $\pm$ 2.3	13.0 $\pm$ 0.9	25.9 $\pm$ 1.0	72.1 $\pm$ 1.2	23.6 $\pm$ 1.0
qwen-3.6-27B-FP8	25.0 $\pm$ 0.9	70.5 $\pm$ 1.1	22.4 $\pm$ 0.9	25.8 $\pm$ 1.0	70.7 $\pm$ 0.9	21.8 $\pm$ 0.8	52.2 $\pm$ 2.2	12.3 $\pm$ 1.0	26.2 $\pm$ 1.0	71.5 $\pm$ 1.1	23.9 $\pm$ 1.0

Table 4: Performance comparison of MedFeat against CAAFE, OCTree, and FeatLLM across different LLM backbones. Results are reported as mean and std over 7 tasks and 2 models. Best results per LLM and metric are shown in bold . FeatLLM does not report F1.

Figure 2: Feature transformation code generated by MedFeat for logistic regression and XGBoost

Figure 3: Feature patterns of the LLM-based feature engineering generated features for two models

Figure 4: Evaluations on IORD data enhanced by transferring MedFeat-generated features from MIMIC

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter	Range [Min, Max]
max_depth	[2, 7]
n_estimators	[100, 2000]
min_child_weight	[10, 100]
max_delta_step	[0, 10]
subsample	[0.5, 1.0]
learning_rate	[0.01, 0.5]

Appendix

Table 5: Hyperparameter search space for XGBoost.

Parameter	Range [Min, Max]
C	[ $10^{-4}$ , $10^{4}$ ]

Appendix

Table 6: Hyperparameter search space for logistic regression.

Dataset	Task	N	$p$	Ratio (+:-)	Task Description
IORD	Mortality-Ward	92,237	31	1:313	Predict whether the patient will die within 1 day.
IORD	Discharge	92,237	31	1:4	Predict whether the patient will be discharged within 1 day.
MIMIC	Mortality-ICU	90,257	16	1:47	Predict whether the patient will die within 24 hours.
	Readmission	74,044	17	1:4	Predict whether the patient will be readmitted within 30 days.
	Heart Failure	74,044	17	1:17	Predict whether the patient will develop heart failure within 30 days.
HRS	Mortality-10yrs	8,899	41	1:9	Predict whether the participant will die within 10 years.

Appendix

Table 7: Statistics of the datasets and prediction tasks.

N

denotes the number of samples,

p

the number of features, and Ratio (+:-) the class balance between positive and negative samples.

Feature	Type
Age	Numerical
Sex	Categorical
Index of multiple deprivation score	Numerical
Admission method code	Categorical
Hours in hospital after admission	Numerical
Number of unique medicines received last 24h	Numerical

Appendix

Table 8: Input features and data types for IORD data.

Feature	Type
Age at ICU admission	Numerical
Sex	Categorical
Admission type	Categorical
Admission location	Categorical
Insurance	Categorical
Language	Categorical

Appendix

Table 9: Input features and MIMIC data types for task Mortality-ICU

Feature	Type
Age at ICU admission	Numerical
Sex	Categorical
Admission type	Categorical
Admission location	Categorical
Insurance	Categorical
Language	Categorical

Appendix

Table 10: Input features and MIMIC data types for task Readmission and Heart Failure

Feature	Type
Respondent gender	Categorical
Respondent race	Categorical
Respondent years of education	Numerical
Respondent veteran status	Binary
Respondent marital status at wave $w$ $(w\in\{8,9,10\})$	Categorical
Respondent age (years) at interview at wave $w$ $(w\in\{8,9,10\})$	Numerical

Appendix

Table 11: Input features and HRS data types

MedFeat vs.	mean $\Delta$	Holm-adj. $p$
Baseline	1.71	0.039*
OpenFE	7.83	0.039*
CAAFE	1.61	0.039*
FeatLLM	20.18	0.039*
OCTree	1.34	0.039*

Appendix

Table 12: One-sided Wilcoxon signed-rank test with Holm correction on AUROC (MedFeat vs. baselines, 7 tasks, learners averaged within task). mean

\Delta

denotes the mean AUROC difference in favor of MedFeat. *:

p<0.05

Method	Cog-Impairment	Mortality-10yrs	Mortality-Ward	Discharge	Mortality-ICU	Heart Failure	Readmission
Baseline	56.3 $\pm$ 1.9	25.3 $\pm$ 0.8	3.6 $\pm$ 0.4	31.3 $\pm$ 0.8	11.6 $\pm$ 1.0	7.4 $\pm$ 0.2	18.8 $\pm$ 0.4
OpenFE	48.3 $\pm$ 3.0	18.6 $\pm$ 5.6	1.7 $\pm$ 0.8	26.2 $\pm$ 1.3	7.1 $\pm$ 4.2	10.7 $\pm$ 8.0	18.2 $\pm$ 0.4
CAAFE	56.3 $\pm$ 1.9	25.4 $\pm$ 0.9	3.6 $\pm$ 0.4	31.3 $\pm$ 0.8	11.8 $\pm$ 1.1	7.6 $\pm$ 0.4	19.0 $\pm$ 0.4
FeatLLM	34.2 $\pm$ 2.6	12.8 $\pm$ 3.0	2.1 $\pm$ 1.3	17.4 $\pm$ 0.2	2.1 $\pm$ 0.0	4.5 $\pm$ 0.0	15.7 $\pm$ 0.0
OCTree	56.3 $\pm$ 1.9	25.0 $\pm$ 0.9	3.4 $\pm$ 0.4	31.3 $\pm$ 0.8	12.8 $\pm$ 1.1	7.6 $\pm$ 0.3	19.0 $\pm$ 0.3
MedFeat	58.8 $\pm$ 1.6	26.3 $\pm$ 1.1	3.9 $\pm$ 0.7	34.6 $\pm$ 0.7	13.6 $\pm$ 1.1	7.8 $\pm$ 0.4	19.9 $\pm$ 0.5

Appendix

Table 13: Per-task AUPRC (%) under logistic regression averaged over 5 random splits. Bold indicates the best method per task.

Method	Cog-Impairment	Mortality-10yrs	Mortality-Ward	Discharge	Mortality-ICU	Heart Failure	Readmission
Baseline	54.9 $\pm$ 1.9	19.6 $\pm$ 2.5	1.7 $\pm$ 0.3	37.0 $\pm$ 0.6	18.1 $\pm$ 2.7	8.2 $\pm$ 0.6	20.3 $\pm$ 0.5
OpenFE	55.2 $\pm$ 1.3	20.8 $\pm$ 2.6	2.6 $\pm$ 1.9	34.8 $\pm$ 0.9	12.1 $\pm$ 0.9	7.4 $\pm$ 0.4	19.7 $\pm$ 0.3
CAAFE	55.1 $\pm$ 1.6	19.7 $\pm$ 2.2	2.2 $\pm$ 1.2	37.1 $\pm$ 0.7	16.2 $\pm$ 2.3	8.3 $\pm$ 0.5	20.6 $\pm$ 0.6
FeatLLM	31.0 $\pm$ 1.1	11.5 $\pm$ 2.0	0.3 $\pm$ 0.0	17.3 $\pm$ 0.0	2.1 $\pm$ 0.0	4.5 $\pm$ 0.0	15.7 $\pm$ 0.0
OCTree	55.5 $\pm$ 1.3	19.1 $\pm$ 2.4	1.9 $\pm$ 0.5	37.0 $\pm$ 0.5	18.0 $\pm$ 2.1	8.3 $\pm$ 0.2	20.4 $\pm$ 0.3
MedFeat	57.2 $\pm$ 0.9	21.7 $\pm$ 1.5	2.4 $\pm$ 0.6	37.4 $\pm$ 0.5	20.3 $\pm$ 1.9	8.9 $\pm$ 0.4	20.9 $\pm$ 0.6

Appendix

Table 14: Per-task AUPRC (%) under XGBoost averaged over 5 random splits. Bold indicates the best method per task.

Figure 5: Total token usage of different methods. Mean and std of the usage are averaged across 7 tasks.

Metric	Baseline	MedFeat	$\Delta$
Deaths caught (TP)	36.0	39.2	+3.2
Deaths missed (FN)	23.0	19.8	–3.2
False alarms (FP)	5225.6	4318.2	–907.4
F1	1.4	1.8	+0.4

Appendix

Table 15: Confusion matrix analysis on Mortality-Ward (XGBoost,

N=18{,}447

test episodes), averaged over 5 random splits at the validation-selected operating point.

Method	Mortality-Ward		Mortality-10yrs		Mortality-ICU
Method	F1	AUROC	F1	AUROC	F1	AUROC
Raw RF	2.6 $\pm$ 0.3	72.2 $\pm$ 3.8	28.1 $\pm$ 2.1	74.0 $\pm$ 2.2	10.9 $\pm$ 0.4	79.4 $\pm$ 1.5
MedFeat + GI	2.7 $\pm$ 0.3	73.8 $\pm$ 3.2	29.4 $\pm$ 3.1	74.6 $\pm$ 2.4	11.4 $\pm$ 1.9	79.7 $\pm$ 1.1
MedFeat + SHAP	2.8 $\pm$ 0.3	74.5 $\pm$ 3.3	29.8 $\pm$ 1.4	74.7 $\pm$ 1.9	11.5 $\pm$ 0.3	79.9 $\pm$ 0.5

Appendix

Table 16: Performance (%) of MedFeat with a Random Forest downstream model under Gini importance (GI) and SHAP attribution. Mean and standard deviation over 5 random splits.

Task	Baseline		MedFeat
Task	F1	AUROC	F1	AUROC
Mortality-ICU	9.6 $\pm$ 1.0	78.4 $\pm$ 2.7	10.4 $\pm$ 1.0	80.3 $\pm$ 1.0
Mortality-10yrs	19.5 $\pm$ 1.3	60.7 $\pm$ 1.8	27.7 $\pm$ 2.5	73.3 $\pm$ 3.7
Heart Failure	13.8 $\pm$ 0.5	65.5 $\pm$ 1.4	13.9 $\pm$ 0.6	66.2 $\pm$ 1.4
Mortality-Ward	1.8 $\pm$ 0.5	72.8 $\pm$ 1.7	2.0 $\pm$ 0.5	79.2 $\pm$ 3.8
Cog-Impairment	58.9 $\pm$ 1.3	76.4 $\pm$ 1.4	59.4 $\pm$ 1.3	77.1 $\pm$ 1.5

Appendix

Table 17: Performance (%) of MedFeat with an MLP downstream learner across the 7 clinical prediction tasks (mean and std over 5 random splits). MedFeat improves over the raw-feature baseline on every task and both metrics.

Task	Acc.	Rej.	Inv.	% Runs $\uparrow$
Cog-Impairment	6.0	4.8	0.1	100%
Mortality-10yrs	4.1	5.3	0.2	90%
Mortality-Ward	3.4	4.2	0.0	80%
Discharge	4.5	4.0	0.3	90%
Mortality-ICU	4.3	3.7	0.2	100%
Heart Failure	6.2	4.1	0.1	100%

Appendix

Table 18: Accepted, rejected, and invalid feature transformations per task, averaged over 5 splits and 2 downstream models. % Runs

\uparrow

denotes the proportion of runs in which test AUROC improves over the raw baseline.

Task	Pattern	Example	Ref.
Mortality-Ward	Clinical scores	Early Warning Score (EWS)	Williams (2022)
Mortality-10yrs	Behavioral risk	Alcohol $\times$ depression	Yan et al. (2025)
Heart Failure	Demographics + vitals	Age $\times$ heart rate change	Ostrowska et al. (2024)
Readmission	Vital sign instability	Var(SpO 2 ) $\times$ Var(respiratory rate)	Nguyen et al. (2017)

Appendix

Table 19: Representative accepted feature patterns, with example transformations and supporting clinical literature.

Method	F1	AUROC	AUPRC
CAAFE	25.1 $\pm$ 0.7	70.6 $\pm$ 1.2	22.4 $\pm$ 1.0
FeatLLM	–	58.9 $\pm$ 4.4	15.1 $\pm$ 1.8
MedFeat	26.2 $\pm$ 1.0	71.5 $\pm$ 1.1	23.9 $\pm$ 1.0

Appendix

Table 20: MedFeat versus the original few-shot implementations of CAAFE and FeatLLM under the Qwen-3.6-27B backbone. Mean

\pm

std over 7 tasks and 2 downstream models.

	Mortality-ICU	Mortality-10yrs	Heart Failure	Mortality-Ward	Cog-Impairment	Discharge	Readmission
# features	21	84	21	28	84	28	21
XGB F1	10.1 $\pm$ 0.5	25.2 $\pm$ 2.1	13.7 $\pm$ 0.4 ‡	1.5 $\pm$ 0.3	56.8 $\pm$ 1.3	40.5 $\pm$ 0.5	29.5 $\pm$ 1.0 †
XGB AUC	76.5 $\pm$ 1.4	68.8 $\pm$ 2.1	65.5 $\pm$ 1.9 †	74.6 $\pm$ 2.7	74.3 $\pm$ 1.4	71.2 $\pm$ 0.6	56.3 $\pm$ 1.0 †
LR F1	9.2 $\pm$ 0.8	27.5 $\pm$ 3.7 †	14.4 $\pm$ 0.2 ‡	2.3 $\pm$ 0.7	57.5 $\pm$ 1.1	38.5 $\pm$ 0.5	30.1 $\pm$ 0.8 ‡
LR AUC	77.7 $\pm$ 1.2	74.9 $\pm$ 1.8 †	68.1 $\pm$ 1.1 ‡	84.0 $\pm$ 1.4	75.7 $\pm$ 1.0	69.6 $\pm$ 0.2	57.4 $\pm$ 0.8 †

Appendix

Table 21: Handcrafted temporal-aggregation baseline with SHAP-based selection (mean and std over 5 random splits, default hyperparameters).

\dagger

: worse than the raw baseline in Table 1 ;

\ddagger

: outperforms MedFeat.

$\beta$	Mortality-Ward	Discharge	Heart Failure	Mean
0.000	+0.52	+0.66	+0.28	+0.49
0.005	+0.38	+0.66	+0.38	+0.47
0.010	+0.40	+0.90	+0.70	+0.67
0.020	+0.28	+0.74	+0.76	+0.59

Appendix

Table 22: Sensitivity of MedFeat to the tolerance threshold

\beta

. Values are absolute F1 improvements (percentage points) over the raw baseline under XGBoost.

Explore similar work

TabPrep: Closing the Feature Engineering Gap in Tabular Benchmarks
Jun 1, 2026Andrej Tschalzev, Nick Erickson, Yuyang Wang +4Tabular Learning Feature Engineering
University of Mannheim · Prior Labs · AWS +3
Adaptive Graph-of-Islands Evolution for Automatic Feature Engineering with LLMs
Jul 25, 2026Sha Li, Naren RamakrishnanFeature Engineering Tabular Learning
Department of Computer Science Virginia Tech Blacksburg, VA 24060 · Department of Computer Science Virginia Tech Alexandria, VA 22305
Medical Heuristic Learning: An LLM-Driven Framework for Interpretable and Auditable Clinical Decision Rules
Jun 15, 2026Wei Xu, Ke Yang, Gang Luo +4Clinical Decision Support Electronic Health Records
The Centre for Artificial Intelligence Driven Drug Discovery, Macao Polytechnic University, Macao SAR · The Key Laboratory of Short-Range Radio Equipment Testing and Evaluation, Ministry of Industry and Information Technology, and the Terahertz Science Application Center (TSAC), Beijing Institute of Technology, Zhuhai, China · The Faculty of Education, The University of Hong Kong, Hong Kong SAR +2

Jun 1, 2026, cs.LGJ/K move · Enter open · S save

TabPrep: Closing the Feature Engineering Gap in Tabular Benchmarks

Andrej Tschalzev, Nick Erickson, Yuyang Wang, Huzefa Rangwala+3

University of Mannheim · Prior Labs · AWS+3

Tabular Learning Feature Engineering Preprocessing

Open paper AskPDF arXiv

Progress in tabular machine learning has largely focused on increasingly sophisticated model architectures. At the same time, feature engineering remains a critical yet underexplored component of real-world modeling pipelines that is entirely absent from modern benchmarks, which creates an unquantified evaluation gap. In this work, we introduce TabPrep, a lightweight preprocessing pipeline composed of feature generators that are carefully designed to target three specific structural data patterns. We show that many widely used model classes exhibit predictable blind spots to these patterns and that systematic feature engineering alone can establish new peak performance. Across the TabArena benchmark, integrating TabPrep into model training and tuning consistently improves performance for tree-based, neural, linear, and foundation models, often surpassing gains achieved by model-centric innovations alone. TabPrep outperforms previous automated feature engineering approaches in performance, efficiency, and applicability across datasets, enabling integration into large-scale benchmarks. By releasing TabPrep (see https://github.com/atschalz/tabprep), we enable researchers to integrate feature engineering into their benchmarking setup, filling a longstanding gap in tabular evaluations.

Abstract

Explore similar work

TabPrep: Closing the Feature Engineering Gap in Tabular Benchmarks

Adaptive Graph-of-Islands Evolution for Automatic Feature Engineering with LLMs

Medical Heuristic Learning: An LLM-Driven Framework for Interpretable and Auditable Clinical Decision Rules