llm-feat 0.2.2

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

llm-feat

Automatically generate feature engineering code for pandas DataFrames using LLMs. Get context-aware, target-specific features that understand your domain.

Installation

pip install llm-feat

Quick Start

import pandas as pd
import llm_feat

llm_feat.set_api_key("your-openai-api-key")  # or set OPENAI_API_KEY env var

# Your data
df = pd.DataFrame({
    'income': [50000, 60000, 70000],
    'expenses': [30000, 35000, 40000],
    'target': [1, 0, 1]
})

# Metadata describing your columns
metadata_df = pd.DataFrame({
    'column_name': ['income', 'expenses', 'target'],
    'description': ['Annual income', 'Annual expenses', 'Binary target'],
    'data_type': ['numeric', 'numeric', 'numeric'],
    'label_definition': [None, None, '1 if positive, 0 if negative']
})

# Generate features
code = llm_feat.generate_features(df, metadata_df, mode='code')
print(code)

Generated Code:

import numpy as np

df['income_to_expense_ratio'] = np.where(df['expenses'] != 0, df['income'] / df['expenses'], np.nan)
df['savings'] = df['income'] - df['expenses']
df['savings_to_income_ratio'] = np.where(df['income'] != 0, df['savings'] / df['income'], np.nan)

Feature Reports

Get detailed explanations of why each feature was generated:

code, report = llm_feat.generate_features(
    df, metadata_df, mode='code', return_report=True
)
print(report)

Example Report:

FEATURE REPORT
==============

1. DOMAIN UNDERSTANDING:
   - Problem: Predicting binary target based on income and expenses
   - Key relationships: Income-to-expense ratios indicate financial health

2. GENERATED FEATURES EXPLANATION:
   - Feature: income_to_expense_ratio
     Rationale: Higher ratios indicate better financial stability
     Domain Relevance: Directly related to predicting positive outcomes

Direct Mode

Add features directly to your DataFrame:

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

Key Features

• Context-aware: Uses column descriptions to generate relevant features
• Target-aware: Generates features specific to your prediction task
• Categorical support: Automatic encoding for categorical columns
• Jupyter integration: Code auto-injected into next cell
• Feature reports: Understand the reasoning behind each feature

Documentation

• Read the Docs - Full documentation
• API Reference - Complete parameter documentation
• Examples - Jupyter notebook examples

Development

git clone https://github.com/codeastra2/llm-feat.git
cd llm-feat
conda create -n llm_feat_310 python=3.10.19 -y
conda activate llm_feat_310
poetry install
poetry run pytest

License

MIT License - see LICENSE file for details.

Author

Srinivas Kumar - @codeastra2

Links

• GitHub
• API Reference
• Changelog