hrboost 0.1.2

HRBoost: Hierarchical Refined Boost - GBDT with Non-monotonic Bayesian Hierarchical Clustering

HRBoost (Hierarchical Refined Boost)

HRBoost is a fast, lightweight Gradient Boosting Decision Tree (GBDT) library built in C++ and Python. It introduces a Non-monotonic Bayesian Hierarchical Clustering (LNM-BHC, $k=3$) algorithm inside its core engine to find optimal splits for high-cardinality categorical variables with zero manual parameter tuning.

HRBoost is 100% compliant with the scikit-learn API, offering both HRBoostClassifier and HRBoostRegressor.

---

Installation

pip install hrboost

---

Hyperparameter Reference

HRBoostClassifier and HRBoostRegressor accept the following parameters in their constructors:

Core GBDT Parameters

• n_estimators (int, default=200): The number of boosting rounds (trees to build).
• learning_rate (float, default=0.1): Shrinkage rate applied to each tree's update to prevent overfitting.
• max_depth (int, default=4): Maximum depth of each decision tree.
• max_leaves (int, default=64): Maximum number of leaves allowed per tree.
• reg_lambda (float, default=1.0): L2 regularization term on weights. It also scales the baseline regularization for Bayesian Hierarchical Clustering.
• subsample (float, default=0.8): Fraction of training samples randomly chosen to train each tree.
• colsample_bytree (float, default=1.0): Fraction of features randomly selected for building each tree.
• n_bins (int, default=32): Maximum number of discrete bins to bucket continuous features.

Split Constraints

• min_child_weight (float, default=0.1): Minimum sum of instance Hessian needed in a child node.
• gamma (float, default=0.0): Minimum loss reduction required to make a split.
• max_delta_step (float, default=0.0): Maximum delta step allowed for each tree's leaf output (useful for highly unbalanced classes).

System & Features

• cat_features (list of int, default=None): List of feature indices to be treated as categorical features.
• random_state (int, default=0): Seed for random number generators (subsampling, colsample).
• verbose (bool, default=True): Controls C++ engine logging during training.

---

Environment Variables for Advanced Tuning

HRBoost exposes internal engine dynamics through system environment variables to avoid hyperparameter inflation:

• COHESION_REG (float, default=0.3):
  • Controls the intensity of the Dynamic Cohesion Regularization during tree splitting.
  • A cohesion penalty factor is computed dynamically based on the difference in predicted leaf values between prospective children. If child leaf predictions diverge excessively, L2 regularization is dynamically increased.
  • Set export COHESION_REG=0.0 to disable this penalty. High-noise categorical settings benefit from higher values (e.g., 0.5 or 1.0).
• MIN_CAT_COUNT (float, default=automatically scaled):
  • The minimum count required for a categorical bin to participate in BHC clustering. It helps filter out extremely rare categorical values.

---

Quick Start

1. Classification

from sklearn.datasets import load_digits
from sklearn.model_selection import train_test_split
from hrboost import HRBoostClassifier

digits = load_digits()
X_train, X_test, y_train, y_test = train_test_split(
    digits.data, digits.target, test_size=0.2, random_state=42
)

clf = HRBoostClassifier(n_estimators=100, learning_rate=0.1, max_depth=4)
clf.fit(X_train, y_train)

print(f"Test Accuracy: {clf.score(X_test, y_test):.4f}")

2. Regression

from sklearn.datasets import load_diabetes
from sklearn.model_selection import train_test_split
from hrboost import HRBoostRegressor

diabetes = load_diabetes()
X_train, X_test, y_train, y_test = train_test_split(
    diabetes.data, diabetes.target, test_size=0.2, random_state=42
)

reg = HRBoostRegressor(n_estimators=150, learning_rate=0.08, max_depth=4)
reg.fit(X_train, y_train)

print(f"Test R2 Score: {reg.score(X_test, y_test):.4f}")