Published Research — PKBoost AI Labs | Jun 2025 – Present
PKBoost is a novel gradient boosting algorithm fusing Shannon entropy with Newton–Raphson optimization, designed specifically for extreme class imbalance and concept drift.
Key Performance Metrics
- Outperforms XGBoost: +17.9% PR-AUC on credit card fraud detection (0.2% minority class).
- Outperforms LightGBM: +10.4% PR-AUC on the same benchmark.
- Drift Resilience: Only 1.8% degradation under covariate shift vs. 31.8% (XGBoost) and 42.5% (LightGBM).
- Speed: 45-second training on 170K samples.
Technical Innovations
- Systems Optimization: Zero-copy architecture, cache-aware data structures (64-byte alignment), and 8x loop unrolling for SIMD.
- Auto-Tuning: Built-in profiler derives optimal hyperparameters, eliminating manual tuning.
- Versatility: Supports binary classification, multi-class (One-vs-Rest), and regression.
Impact & Adoption
- Downloads: 4,400+ PyPI downloads.
- Research: Published DOI 10.5281/zenodo.17541137.
- Community: 60+ GitHub stars, featured on Kaggle.