Crate gbt_quantile 

gbt-quantile

Pure-Rust gradient-boosted trees with quantile regression.

This crate provides a complete GBT implementation with:

• Squared-error (L2) loss for mean regression
• Pinball loss for quantile regression (P10, P25, P50, P75, P90, etc.)
• Early stopping with periodic validation checks
• Quantile ensembles that train multiple models and enforce monotonicity
• JSON serialization for model persistence and portability

Quick start

use gbt_quantile::{GBTConfig, QuantileEnsemble};

// Generate some data
let x: Vec<Vec<f64>> = (0..100).map(|i| vec![i as f64]).collect();
let y: Vec<f64> = x.iter().map(|r| 2.0 * r[0] + 1.0).collect();

// Train a quantile ensemble
let config = GBTConfig { n_trees: 50, ..GBTConfig::default() };
let quantiles = vec![0.1, 0.5, 0.9];
let ensemble = QuantileEnsemble::train(&x, &y, &quantiles, &config, None);

// Predict
let pred = ensemble.predict(&[50.0]);
println!("P10={}, P50={}, P90={}",
    pred.p(0.1).unwrap(),
    pred.median(),
    pred.p(0.9).unwrap(),
);

Single model

use gbt_quantile::{GBTConfig, trainer};

let x = vec![vec![1.0], vec![2.0], vec![3.0], vec![4.0], vec![5.0]];
let y = vec![2.0, 4.0, 6.0, 8.0, 10.0];

let config = GBTConfig { n_trees: 50, min_samples_leaf: 1, ..GBTConfig::default() };
let model = trainer::train(&x, &y, &config, None);
let prediction = model.predict(&[3.0]);

Re-exports

pub use config::GBTConfig;

pub use ensemble::QuantileEnsemble;

pub use ensemble::QuantilePrediction;

pub use metrics::Metrics;

pub use tree::GradientBoostedTree;

Modules

config

ensemble

Quantile ensemble: train multiple GBT models at different quantiles and predict with monotonicity enforcement.

metrics

Evaluation metrics for regression and quantile models.

split

Data splitting utilities for train/test partitioning.

trainer

Core gradient-boosted tree training algorithm.

tree

Tree model data structures: nodes, leaves, and the full GBT model.