PKBoost is an open-source gradient boosting library optimized for imbalanced classification and concept drift detection. It uses Shannon entropy to guide tree splits and automatically adapts to changing data distributions.

How does PKBoost handle imbalanced data?

PKBoost uses Shannon entropy-guided splitting to find informative patterns in minority classes. It achieves 87.8% PR-AUC on fraud detection with 0.2% positive class, outperforming XGBoost (74.5%) and LightGBM (79.3%).

What is concept drift adaptation?

PKBoost automatically detects when model performance degrades due to distribution shifts and triggers metamorphosis - pruning obsolete trees and growing new ones on recent data. This happens in 2 seconds without manual intervention.

How do I install PKBoost?

Install via pip: 'pip install pkboost' for Python, or 'cargo add pkboost' for Rust. PKBoost has a scikit-learn compatible interface for easy integration.

When should I use PKBoost vs XGBoost?

Use PKBoost for: extreme class imbalance (<5% minority), streaming data with concept drift, or when you need automatic hyperparameter tuning. Use XGBoost for: balanced datasets or when you have time for manual tuning.

PKBOOST

Adaptive Gradient Boosting for a Changing World

Built in Rust. Tuned by Shannon. Designed for Drift.

> GET STARTED > PYTHON PACKAGE

> BENCHMARKS > MATHEMATICS > SOURCE CODE

Quick Start

python

from pkboost import PKBoostClassifier

model = PKBoostClassifier.auto()
model.fit(X_train, y_train)
predictions = model.predict_proba(X_test)

Benchmark Results

Credit Card Fraud (0.2% positive class)

PKBoost 87.8% PR-AUC

LightGBM 79.3% PR-AUC

XGBoost 74.5% PR-AUC

PKBoost: 87.8% • LightGBM: 79.3% • XGBoost: 74.5%

87.8%

PR-AUC on Fraud Detection

Adaptation Time

Auto

Parameter Tuning

Design Goals

Traditional Limitations

• Manual retraining on drift
• Poor minority class recall
• Extensive hyperparameter tuning
• Performance degradation under drift

PKBoost Approach

• Automatic adaptation via metamorphosis
• Shannon entropy for rare patterns
• Auto-tuned from data statistics
• Stable performance under drift

Concept Drift

Adapts to distribution shifts without manual intervention through automatic metamorphosis.

Class Imbalance

Shannon entropy guidance for improved minority class detection in imbalanced datasets.

Auto-Tuning

Hyperparameters automatically configured based on dataset characteristics.

Use Cases

Fraud Detection

Credit cards, insurance claims, transactions

Medical Diagnosis

Rare disease detection, risk assessment

Cybersecurity

Intrusion detection, malware classification

Anomaly Detection

Equipment failure, quality control