Evolutionary Feature Engineering (EFE)

Code for the paper Evolutionary Feature Engineering for Structured Data.

EFE uses LLM-guided evolutionary search to discover feature-engineering programs for structured data. The downstream model stays fixed. EFE only changes the preprocessing transformation placed before it.

• EFE-Time evolves invertible normalizations for time-series forecasting.
• EFE-Tab evolves compact feature generator programs for tabular classification.

The includes the EFE pipelines and the runners for evolving, validating, and scoring candidate programs.

Installation

From the repository root:

python -m venv .venv
source .venv/bin/activate

pip install -e .
pip install -r requirements.txt

The default configs use AWS Bedrock. Set your LLM credentials before running:

export AWS_BEARER_TOKEN_BEDROCK="your-token"

OpenAI-compatible endpoints can also be configured in the relevant config.yaml file.

Quick Start

EFE-Time

export GIFT_EVAL=/path/to/gift_eval_data
export TS_DATASET=covid_deaths

python openevolve-run.py \
  EFE-Time/initial_program.py \
  EFE-Time/evaluator.py \
  --config EFE-Time/config.yaml \
  --iterations 100 \
  --output EFE-Time/openevolve_output

EFE-Tab

export TABARENA_TASK=churn

python openevolve-run.py \
  EFE-Tab/initial_program.py \
  EFE-Tab/evaluator.py \
  --config EFE-Tab/config.yaml \
  --iterations 100 \
  --output EFE-Tab/openevolve_output

Outputs are written to the selected openevolve_output/ directory.

Repository Structure

EFE/
├── EFE-Time/          # time-series transformation evolution
├── EFE-Tab/           # tabular feature-program evolution
├── openevolve/        # bundled evolutionary program-search framework
├── openevolve-run.py  # command-line entry point
├── requirements.txt
└── INSTALL.md

Method-specific setup details are available in EFE-Time/README.md, EFE-Tab/README.md, and INSTALL.md.

Citation

If you use this repository, please cite:

@misc{taga2026evolutionaryfeatureengineeringstructured,
      title={Evolutionary Feature Engineering for Structured Data}, 
      author={Ege Onur Taga and Yilin Zhuang and M. Emrullah Ildiz and Petros Mol and Abhimanyu Das and Karthik Duraisamy and Samet Oymak},
      year={2026},
      eprint={2607.01548},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2607.01548}, 
}

Acknowledgments

This repository builds on OpenEvolve as its evolutionary program-search framework.

License

This project is released under the Apache-2.0 license.