llm-feat 0.2.0

Automated feature engineering using Large Language Models (LLMs) for tabular data

llm-feat

llm-feat is a Python library that automatically generates feature engineering code for pandas datasets using Large Language Models. Generate context-aware, target-specific features that can be automatically added to your DataFrame or customized before execution.

🌟 Key Features

• 🤖 LLM-Powered: Uses OpenAI models to generate context aware feature engineering code.
• 📝 Dual Modes: Get code suggestions or directly add features
• 🔧 Jupyter Support: Automatically injects code into next cell in Jupyter notebooks
• 🎯 Metadata-Driven: Uses column descriptions to generate contextually relevant features
• 🎯 Target-Aware: Generates features specifically relevant to your prediction task

📦 Installation

To install the latest release of llm-feat from PyPI:

pip install llm-feat

🚀 Quick Start

1. Setup

import pandas as pd
import llm_feat

# Set your OpenAI API key
llm_feat.set_api_key("your-openai-api-key-here")
# Or use environment variable: export OPENAI_API_KEY="your-key"

2. Prepare Data and Metadata

Consider the problem of healthy and unhealthy labelling:

# Your dataset with target column
df = pd.DataFrame({
    'height': [170, 175, 180, 165, 185, 172, 178, 168, 182, 174],
    'weight': [70, 75, 80, 65, 85, 72, 78, 68, 83, 74],
    'bmi': [24.2, 24.5, 24.7, 23.9, 24.8, 24.3, 24.6, 24.1, 25.0, 24.4],
    'health_score': [1, 1, 0, 1, 0, 1, 1, 1, 0, 1]  # Target: 1=healthy, 0=unhealthy
})

# Metadata with column descriptions and target definition
metadata_df = pd.DataFrame({
    'column_name': ['height', 'weight', 'bmi', 'health_score'],
    'description': [
        'Height in centimeters',
        'Weight in kilograms',
        'Body Mass Index',
        'Health classification score'
    ],
    'data_type': ['numeric', 'numeric', 'numeric', 'numeric'],
    'label_definition': [None, None, None, '1 if healthy, 0 if unhealthy']
})

3. Generate Features

Mode 1: Get Code (Recommended for Jupyter)

We will now generate features for the above dataset based on feature descriptions in the metadata and prior knowledge of the LLM.

code = llm_feat.generate_features(df, metadata_df, mode='code', model='gpt-4o-mini')
print(code)

Example Output:

import numpy as np

df['height_weight_ratio'] = df['height'] / df['weight'].replace(0, np.nan)
df['bmi_squared'] = df['bmi'] ** 2
df['weight_bmi_interaction'] = df['weight'] * df['bmi']
df['health_score_bmi_diff'] = df['health_score'] - df['bmi'].apply(lambda x: 1 if x < 24.9 else 0)
df['height_bmi_category'] = pd.cut(df['bmi'], bins=[0, 18.5, 24.9, 29.9, np.inf], 
                                    labels=['Underweight', 'Normal', 'Overweight', 'Obese'])

Note: In Jupyter notebooks, the code is automatically injected into the next cell. The LLM generates target-aware features that are relevant to predicting health_score.

Mode 2: Direct Feature Addition

df_with_features = llm_feat.generate_features(df, metadata_df, mode='direct', model='gpt-4o-mini')
print(df_with_features.head())

Example Output:

   height  weight   bmi  health_score  height_weight_ratio  bmi_squared  \
0     170      70  24.2             1             2.428571      585.64   
1     175      75  24.5             1             2.333333      600.25   
2     180      80  24.7             0             2.250000      610.09   

   weight_bmi_interaction  health_score_bmi_diff height_bmi_category  
0                1694.0                      0              Normal  
1                1837.5                      0              Normal  
2                1976.0                      0              Normal  

📖 Usage Examples

Jupyter Notebook Example

See example_llm_feat.ipynb for complete usage examples in Jupyter notebook format.

Python Script Example

You can also run the example as a standalone Python script:

poetry run python example_llm_feat.py

Note: The notebook and script examples use df as the DataFrame variable name, which is required for the generated code to work directly.

Development Installation

Prerequisites

• Python 3.10.19 or higher (tested with Python 3.10.19)
• Poetry (for development)
• Conda (optional, for environment management)

1. Clone the repository:

   git clone https://github.com/codeastra2/llm-feat.git
   cd llm-feat
   

2. Create conda environment:

   conda create -n llm_feat_310 python=3.10.19 -y;
   conda activate llm_feat_310
   

3. Install Poetry (if not already installed):

   curl -sSL https://install.python-poetry.org | python3 -
   

4. Install dependencies:

   poetry install
   

5. Add env to jupyter kernels(to be used in ipython notebooks):

   poetry run python -m ipykernel install --user --name llm_feat_310 --display-name "Python (llm_feat_310)"
   

Running Tests

poetry run pytest

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🤝 Contributing

Contributions are welcome! Please check our GitHub repository for guidelines.

👤 Author

Srinivas Kumar - [e-mail]

📚 Documentation

• CHANGELOG.md - Version history and changes For any questions or issues, please open an issue on our GitHub repository.