borutashap-modern

A modernized fork of BorutaShap that works with current versions of NumPy 2.0+, SciPy, and scikit-learn. This fork includes performance improvements and bug fixes for SHAP-based feature selection.

Installation

# Install from PyPI (recommended)
pip install borutashap-modern

# With LightGBM support (recommended for speed)
pip install borutashap-modern[lightgbm]

# With all optional dependencies
pip install borutashap-modern[all]

Key Improvements

Compatibility Fixes

• NumPy 2.0+ support: Fixed deprecated np.NaN to np.nan
• SciPy 1.11+ support: Updated binom_test to binomtest with backward compatibility
• Python 3.12+ support: Requires Python 3.12 or higher

Bug Fixes

• RandomForest + SHAP: Fixed 3D array handling and indexing issues
• RandomForest + Gini: Fixed premature feature_importances_ check
• Missing imports: Added required imports (inspect, defaultdict)

Performance Insights

Based on extensive benchmarking:

• LightGBM: Best overall performer (0.6s avg SHAP time, F1=0.875)
• XGBoost: Good balance (1.6s avg SHAP time, F1=0.868)
• RandomForest: Best F1 on small datasets (F1=0.935 @ 1k samples)
• GradientBoosting: Highest accuracy but slow (13s avg SHAP time)

Requirements

• Python 3.12+
• NumPy 2.0+
• pandas 2.0+
• scikit-learn 1.3+
• SHAP 0.45+
• LightGBM 4.0+ (optional, recommended)
• XGBoost 2.0+ (optional)

Quick Start

from BorutaShap import BorutaShap
from lightgbm import LGBMClassifier
from sklearn.datasets import make_classification
import pandas as pd

# Generate sample data
X, y = make_classification(n_samples=1000, n_features=20, n_informative=10)
X = pd.DataFrame(X, columns=[f'feature_{i}' for i in range(20)])

# Initialize with LightGBM (recommended for speed)
model = LGBMClassifier(n_estimators=50, max_depth=5, verbose=-1)

# Run BorutaShap
fs = BorutaShap(
    model=model,
    importance_measure='shap',  # or 'gini' for tree-based models
    classification=True
)

fs.fit(X=X, y=y, n_trials=100, random_state=42)

# Get results
print(f"Accepted features: {fs.accepted}")
print(f"Rejected features: {fs.rejected}")
print(f"Tentative features: {fs.tentative}")

Performance Recommendations

Model Selection Guide

Use Case	Recommended Model	F1 Score	SHAP Speed
Small data (<5k samples)	RandomForest	0.935	0.15s
Medium data (5-50k)	LightGBM	0.90	0.5-2s
Large data (>50k)	LightGBM	0.89	2-5s
Best accuracy	GradientBoosting	0.91	10-50s
Production/speed critical	LightGBM	0.88	<2s

Dataset Size Impact

• Samples: More samples → better F1 (all models improve 5-9%)
• Features: More features → worse F1 (especially RandomForest: -15% from 10→200 features)
• Sweet spot: 5-10k samples with ≤50 features

Feature Importance Methods

• SHAP: More accurate but ~11x slower than Gini
• Gini: Fast but only for tree-based models (not XGBoost)
• Recommendation: Use SHAP for final models, Gini for exploration

Supported Models

✅ Fully Supported:

• LightGBM (fastest SHAP)
• XGBoost (SHAP only)
• RandomForest (both SHAP and Gini)
• ExtraTrees (both SHAP and Gini)
• GradientBoosting (both SHAP and Gini)

❌ Not Supported:

• BaggingClassifier (SHAP TreeExplainer incompatible)
• SVM, Neural Networks (no tree structure)

Testing

# Run basic test
python examples/test_basic.py

# Run performance comparison
python examples/compare_models.py

# Test with your data
python examples/test_custom.py --data your_data.csv

Changes from Original

1. Fixed NumPy 2.0 compatibility (src/BorutaShap.py:L384-394)
2. Fixed SciPy binomial test import (src/BorutaShap.py:L8-13)
3. Fixed RandomForest SHAP 3D array handling (src/BorutaShap.py:L250-260)
4. Fixed RandomForest Gini importance check (src/BorutaShap.py:L150-155)
5. Added Python 3.12 support (setup.py)
6. Added comprehensive benchmarks (examples/benchmark.py)

Citation

If you use this fork, please cite both the original and this fork:

# Original BorutaShap
@software{boruta_shap,
  author = {Eoghan Keany},
  title = {BorutaShap: A wrapper feature selection method using Boruta and SHAP},
  url = {https://github.com/Ekeany/Boruta-Shap},
  year = {2020}
}

# This fork
@software{boruta_shap_modern,
  author = {BlackArbsCEO},
  title = {BorutaShap Modern Fork: Compatible with NumPy 2.0+},
  url = {https://github.com/BlackArbsCEO/Boruta-Shap},
  year = {2024}
}

Contributing

Contributions are welcome! Please:

1. Fork the repository
2. Create a feature branch
3. Run tests with Python 3.9+
4. Submit a pull request

License

MIT License (same as original)

Acknowledgments

• Original author: Eoghan Keany
• SHAP library: lundberg/shap
• Boruta algorithm: Boruta R package