"""

TreeBoost - High-performance Gradient Boosted Decision Tree engine

A pure Rust GBDT implementation designed for large-scale tabular data

with robust handling of dirty/noisy data.

Key Features:

- Histogram-based training: u8 bins for memory efficiency

- Shannon Entropy regularized splits: Drift-resilient objective

- Pseudo-Huber loss: Robust to outliers

- Split Conformal Prediction: Distribution-free prediction intervals

- Zero-copy serialization: Fast model loading via rkyv

Usage:

import numpy as np

from treeboost import GBDTConfig, GBDTModel

# Prepare data

X = np.random.randn(1000, 10).astype(np.float32)

y = (X[:, 0] + X[:, 1] * 2 + np.random.randn(1000) * 0.1).astype(np.float32)

# Configure model

config = GBDTConfig()

config.num_rounds = 100

config.max_depth = 6

config.learning_rate = 0.1

# Train

model = GBDTModel.train(X, y, config)

# Predict

predictions = model.predict(X)

# Feature importance

importances = model.feature_importance()

# Save/load

model.save("model.rkyv")

loaded = GBDTModel.load("model.rkyv")

Advanced Features:

# Pseudo-Huber loss (robust to outliers)

config.use_pseudo_huber_loss(delta=1.0)

# Conformal prediction intervals

config.calibration_ratio = 0.2

config.conformal_quantile = 0.9

preds, lower, upper = model.predict_with_intervals(X)

# Early stopping

config.early_stopping_rounds = 10

config.validation_ratio = 0.2

# Subsampling (stochastic gradient boosting)

config.subsample = 0.8 # Row subsampling

config.colsample = 0.8 # Column subsampling

# Monotonic constraints

# 1 = increasing, -1 = decreasing, 0 = none

config.set_monotonic_constraints([1, -1, 0, 0, 0])

# Feature interaction constraints

# Features in same group can interact together

config.set_interaction_groups([[0, 1, 2], [3, 4]])

"""

from ._core import GBDTConfig, GBDTModel

__all__ = ["GBDTConfig", "GBDTModel"]

__version__ = "0.1.0"