BSSCL-GBM

Hybrid Histogram Gradient Boosting Machine

BSSCL-GBM is a highly memory-efficient, pure-Python implementation of Gradient Boosting Trees optimized with Numba. It achieves mathematical and statistical parity with C++ giants like XGBoost and LightGBM while maintaining a fundamentally lighter memory footprint.

Developed by Bhaskar, BSSCL-GBM introduces novel algorithmic paradigms including Hybrid uint8 Histogram Binning and mathematically dampened balanced_sqrt Class Weighting (power=0.35) specifically engineered for highly imbalanced datasets (e.g., fraud detection, rare disease prediction).

⚡ Features

• Hybrid uint8 Binning: Radically reduces Peak RAM (Resident Set Size) consumption by packing numerical feature splits into highly optimized 8-bit integer arrays.
• Native Missing Value Handling: Automatically routes NaN values during the histogram construction phase without requiring external imputation, successfully stress-tested up to 70% data corruption.
• Imbalance Eradication: Introduces the balanced_sqrt auto-weighting mechanism (power=0.35) to perfectly dampen severe class imbalances (e.g., 99.9:0.1) without over-penalizing the majority class.
• Noise Resistance: Proven superior performance on datasets with label noise (5% flip rate), outperforming XGBoost, LightGBM, and CatBoost on all metrics.
• Numba JIT Compilation: Achieves near C-level looping speeds while remaining a 100% pure Python package.
• Scikit-Learn API: Fully compatible with standard fit(), predict(), predict_proba() pipelines.
• Zero Data Leakage: Code audit verified — bin edges, class weights, and categorical encoding are computed strictly from training data.

📦 Installation

Install BSSCL-GBM via pip:

pip install bsscl-gbm

🚀 Quickstart

BSSCL-GBM operates exactly like any standard Scikit-Learn estimator.

from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
from bsscl_gbm import HybridHistGBMNumbaV2

# Load Data
X, y = load_breast_cancer(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

# Initialize and Train
model = HybridHistGBMNumbaV2(n_estimators=100, max_depth=6)
model.fit(X_train, y_train)

# Predict
preds = model.predict(X_test)
probs = model.predict_proba(X_test)

For highly imbalanced data (e.g., 99:1 split), simply enable the proprietary dampening mechanism:

model = HybridHistGBMNumbaV2(class_weight='balanced_sqrt')

📊 Scientific Benchmarks (v1.0.1)

BSSCL-GBM has been rigorously benchmarked against XGBoost, LightGBM, and CatBoost across 8 diverse datasets using a strict anti-leakage protocol (stratified split, scaler fitted on train only, 3 random seeds averaged). Every competitor was given its best available auto-weighting for maximum fairness.

1. Extreme Imbalance — Credit Card Fraud (284K rows, 99.8% / 0.2%)

2. Noisy Labels — Overlapping Features (50K rows, 98/2, 5% label noise)

BSSCL-GBM wins all 6 metrics on noisy data — the power=0.35 dampening is exceptionally noise-resistant.

3. High-Dimensional Sparse (100K rows, 100 features, 99/1)

4. Memory Efficiency (Peak RSS)

🔒 Code Audit (v1.0.1)

A comprehensive line-by-line audit of the full 4,299-line source code confirmed:

📁 Repository Structure

• /src/bsscl_gbm: The core pure-Python algorithms.
• /tests: Comprehensive Pytest unit validation suite.
• /benchmarks/academic: A 20-file scientific proving ground covering Ablation, Imbalance Severity, Learning Curves, and Statistical Significance.
• /examples: Jupyter notebooks for immediate onboarding.

🤝 Contributing

We welcome contributions! Please check CONTRIBUTING.md and read our CODE_OF_CONDUCT.md.

📝 License & Academic Citation

Released under the Apache 2.0 License.

If you utilize BSSCL-GBM for academic research, please cite it using the provided CITATION.cff file or the following BibTeX:

@software{bsscl_gbm_2026,
  author = {Bhaskar},
  title = {BSSCL-GBM: Hybrid Histogram Gradient Boosting Machine},
  year = {2026},
  url = {https://github.com/Bhaskar19chowdary/bsscl-gbm}
}