PromptFE

The source code for the paper PromptFE: Automated Feature Engineering by Prompting.

PromptFE uses large language models to automatically suggest and evaluate engineered features for tabular machine learning datasets. It iteratively prompts an LLM to generate feature transformations, evaluates them using cross-validation, and selects the best-performing features.

Requirements

• Python 3.8+
• OpenAI API key

Install dependencies:

pip install -r requirements.txt

Setup

Create a .env file in the project root with your OpenAI API key:

OPENAI_API_KEY=your_api_key_here

Usage

1. To create and evaluate features, run:

python run.py

This prompts GPT to suggest feature transformations, evaluates them on the specified model and dataset, and saves results to the results/ directory.

2. To read and analyze results, run:

python read.py

This loads saved results from results/ and writes a performance summary to log_result.txt.

Project Structure

PromptFE/
├── data/                       # Dataset files
│   ├── *.csv                   # Raw data
│   ├── *.json                  # Dataset metadata (task type, feature info)
│   └── *.txt                   # Feature descriptions
├── dataset_descriptions/       # Extended dataset descriptions
├── params/                     # Model hyperparameter configurations
│   └── {MODEL}-{DATASET}.txt
├── utils/
│   └── tools.py                # Logging and timing utilities
├── run.py                      # Main script: feature generation and evaluation
├── read.py                     # Results analysis script
├── env.py                      # Model training and evaluation environment
├── feat_tree.py                # Feature tree construction and RPN parsing
├── dataset.py                  # Dataset loading and preprocessing
├── search_space.py             # Feature operation definitions
├── metrics.py                  # Custom evaluation metrics
└── requirements.txt

Example Datasets

Dataset	Task
AIDS	Classification
Airfoil	Regression
Bikeshare	Regression
Credit Default	Classification
German Credit	Classification
Housing	Regression
Wine Quality (Red)	Regression

Supported Models

Classification: LogisticRegression, RandomForestClassifier, XGBClassifier, LGBMClassifier

Regression: Lasso, RandomForestRegressor, XGBRegressor, LGBMRegressor

Feature Operations

PromptFE supports the following feature transformation types:

• Unary: log, sqrt, reciprocal, min-max
• Binary: +, -, *, /, mod

Features are represented as trees using canonical Reverse Polish Notation (cRPN) and evaluated iteratively to find the best-performing subset.

Citation

If you use this code, please cite the paper:

@inproceedings{zou-etal-2026-promptfe,
    title = "{P}rompt{FE}: Automated Feature Engineering by Prompting",
    author = "Zou, Yufeng  and  Utke, Jean  and  Klabjan, Diego  and  Liu, Han",
    booktitle = "Proceedings of the 19th Conference of the {E}uropean Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
    month = mar,
    year = "2026",
    address = "Rabat, Morocco",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.eacl-long.28/",
    doi = "10.18653/v1/2026.eacl-long.28",
    pages = "653--681",
    ISBN = "979-8-89176-380-7"
}

A portion of the code is adapted from DIFER.