TowerGB

TowerGB is a fast, enterprise-ready machine learning model for tabular classification. It features a dual-tower ensemble architecture with built-in probability calibration, feature importance attribution, class balancing, and sub-millisecond inference.

How It Works

TowerGB works in four steps:

1. Tower 1 (Accuracy): Trains bootstrap passes and evaluates accuracy.
2. Tower 2 (Risk & Calibration): Evaluates log-loss, Brier score, and sample-wise loss variance.
3. Pareto Arbiter: Computes optimal ensemble voting weights based on accuracy and risk.
4. Temperature Calibration: Minimizes Expected Calibration Error (ECE) via Golden Section Search.

Key Features

• Zero Heavy Dependencies: Pure NumPy and Scikit-Learn.
• Sub-Millisecond Latency: 0.09 ms per batch inference via single collapsed matrix multiplication.
• Feature Importances & Coefficients: Full interpretability via .feature_importances_, .coef_, and .intercept_.
• Imbalanced Data Ready: Built-in class_weight='balanced' and sample_weight support.
• L2 Regularization: Built-in weight decay (l2_reg) to prevent overfitting.
• Parallel Training: Native multi-core CPU scaling (n_jobs=-1).
• Scikit-Learn Standard: 100% compliant with Pipeline, GridSearchCV, cross_val_score, and serialization.

Installation

# Clone the repository
git clone https://github.com/anishupr47-git/TableGB.git
cd TableGB

# Create a virtual environment
python -m venv .venv

# Activate on Windows:
.venv\Scripts\activate

# Or activate on macOS/Linux:
# source .venv/bin/activate

# Install the package
pip install -e ".[test]"

Quick Example

from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
from towergb import TowerGBClassifier

# 1. Create example table data
X, y = make_classification(n_samples=1000, n_features=20, n_classes=3,
                           n_informative=10, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

# 2. Create and train the model with L2 regularization and balanced weights
clf = TowerGBClassifier(n_passes=5, learning_rate=0.1, l2_reg=1e-4, random_state=42)
clf.fit(X_train, y_train)

# 3. Check accuracy
print(f"Accuracy: {clf.score(X_test, y_test):.4f}")

# 4. View feature importance ranking
print(f"Top feature importance: {clf.feature_importances_[:5]}")

# 5. Tune confidence probabilities
clf.calibrate(X_test, y_test)
proba = clf.predict_proba(X_test)
print(f"Calibrated probabilities shape: {proba.shape}")

Settings and Options

TowerGBClassifier Parameters

Public Attributes

Main Methods

Running Tests

# Run all tests
pytest tests/ -v

# Run with test coverage
pytest tests/ -v --cov=towergb --cov-report=term-missing

Running Benchmarks

# Install benchmark tools
pip install -e ".[benchmark]"

# Run speed and accuracy comparison
python benchmarks/run_benchmarks.py

License

MIT