PKBoost: Adaptive Gradient Boosting with Shannon Entropy Guidance and Metamorphic Drift Recovery

PKBoost is a gradient boosting implementation tailored for extremely imbalanced and non-stationary data settings, e.g., fraud detection and anomaly monitoring. The approach proposes to enhance by two essential innovations: (i) Adaptive Entropy Splitting (AES), a criterion based on the Newton- Raphson second order optimization combined with the Shannon Entropy principle in order to more effectively separate minority-class structures, and ii) Hierarchical Adaptation Boosting (HAB): a metamorphic update strategy aimed at observing changes in concept drift by monitoring classifier vulnerabilities and retraining only affected portions of the feature space.

PKBoost can stably keep PR-AUC performance under drift and has a much stronger ability to recall rare events than XGBoost, LightGBM. It is implemented in Rust for speed and accessibility via Python bindings for easy inclusion in data science pipelines.

This release includes the complete open-source code as well as benchmarking scripts, mathematical derivations, and experimental results achieving competitive performance on both Credit Card Fraud data set and a variety of drift scenarios.