SnapBoost

SnapBoost is an instance of a Heterogeneous Newton Boosting Machine (HNBM) — a generalized gradient boosting framework that supports the use of various types of learners aside from trees. Snapboost is an HNBM that mixes decision trees and kernel ridge regressors instead of trees alone. The core HNBM framework is provided by the hnbm package; SnapBoost is a concrete implementation built on top of it.

Unlike XGBoost and LightGBM, which rely exclusively on decision trees as base learners, SnapBoost stochastically selects from a heterogeneous pool of learners at each boosting iteration. This lets the model capture both local, axis-aligned structure (trees) and smooth, global patterns (RBF kernel ridge).

This package is a Python/scikit-learn reimplementation inspired by SnapBoost: A Heterogeneous Boosting Machine (Parnell et al., NeurIPS 2020). See REFERENCES.md for papers, related work, and citation details.

Table of Contents

• Documentation
• Features
• Mathematical Overview
• Installation
• Quick Start
• Examples & Results
• API Reference
  • SnapBoostClassifier / SnapBoostRegressor
  • SnapBoost
  • HNBM
• Parameters
• Docker
• Development
• References & Citation
• License

Documentation

The SnapBoost API documentation lives in docs/. To build locally:

pip install -r docs/requirements.txt
cd docs && make html
# open _build/html/index.html

Documentation is published at https://snapboost.qiancapital.com/ (GitHub Pages). The live docs on / track master (latest). Release snapshots are under /vX.Y.Z/ (for example /v1.0.0/) and are rebuilt from docs/versions.json on each master deploy. Use the version dropdown under SnapBoost in the sidebar to switch between them.

Features

Heterogeneous Gradient Boosting

SnapBoost builds an additive predictor from heterogeneous learners:

$$ F_M(x)=F_0+\sum_{m=1}^{M}\eta_m f_m(x), $$

where $F_0$ is a constant initial prediction, $f_m$ is a decision tree, random-Fourier-feature (RFF) ridge model, or optional linear ridge model, and $\eta_m$ is the learning rate (or a per-round step selected by line search).

At round $m$, let $F_{m-1}(x_i)$ be the current raw prediction and let $\ell(y_i,F)$ be the objective. SnapBoost computes

$$ g_i=\left.\frac{\partial\ell(y_i,F)}{\partial F}\right|{F=F{m-1}(x_i)}, \qquad h_i=\left.\frac{\partial^2\ell(y_i,F)}{\partial F^2}\right|{F=F{m-1}(x_i)}. $$

A second-order Taylor expansion turns the next functional step into weighted least squares. The selected learner is therefore fit to the Newton working response

$$ r_i=-\frac{g_i}{h_i}, \qquad f_m\approx\arg\min_{f\in\mathcal H_{k_m}} \sum_{i=1}^{n} w_i h_i\bigl(r_i-f(x_i)\bigr)^2, $$

where $w_i$ is the observation weight. With the default random strategy, the learner family $k_m$ is sampled from the configured pool: tree depths share probability p_tree, the optional linear learner has probability p_linear, and RFF kernel candidates share the remainder. The update is

$$ F_m(x)=F_{m-1}(x)+\eta_m f_m(x). $$

For squared-error regression, $g_i=2(F-y_i)$ and $h_i=2$, so $r_i=y_i-F$: ordinary residual boosting is recovered. For binary classification SnapBoost encodes labels as $y_i\in{-1,+1}$ and uses logistic loss,

$$ \ell(y,F)=\log(1+e^{-yF}),\quad g=-y,\sigma(-yF),\quad h=\sigma(yF)\sigma(-yF), $$

with class probability $P(y=+1\mid x)=\sigma(F_M(x))$.

The smooth branch approximates a stationary kernel with random features. For the default RBF kernel $k(x,x')=\exp(-\gamma\lVert x-x'\rVert_2^2)$,

$$ \phi_j(x)=\sqrt{\frac{2}{D}}\cos(\omega_j^\top x+b_j), \quad \omega_j\sim\mathcal N(0,2\gamma I), \quad b_j\sim\mathrm{Uniform}(0,2\pi), $$

and weighted ridge regression solves

$$ \hat\beta=\arg\min_\beta \sum_i w_i h_i\bigl(r_i-\phi(x_i)^\top\beta\bigr)^2 +\alpha\lVert\beta\rVert_2^2. $$

Thus, trees model local axis-aligned interactions while RFF ridge learners add smooth global corrections. XGBoost applies a related second-order expansion but restricts every round to a regularized tree and optimizes leaf weights and split gains analytically; SnapBoost instead projects the Newton step onto a randomly selected (or greedily selected) heterogeneous hypothesis class.

See MATH.md for the full derivation, initialization and objective formulas, RFF and tree subproblems, optional training behavior, and a detailed comparison with gradient boosting, Newton tree boosting, and XGBoost.

Installation

From PyPI (recommended):

pip install snapboost

From source:

git clone https://github.com/qiancapital/snapboost.git
cd snapboost
pip install .

Requirements: Python ≥ 3.9, NumPy, scikit-learn, tqdm, hnbm ≥ 1.0.0.

Quick Start

Classification

from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
from snapboost import SnapBoostClassifier

X, y = load_breast_cancer(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42)

model = SnapBoostClassifier(
    num_iterations=100,
    learning_rate=0.1,
    random_state=42,
)
model.fit(X_train, y_train)

print("Accuracy:", model.score(X_test, y_test))
print("Probabilities shape:", model.predict_proba(X_test).shape)  # (n_samples, 2)
model.evaluate(X_test, y_test)  # prints log loss

Regression

from sklearn.datasets import load_diabetes
from sklearn.model_selection import train_test_split
from snapboost import SnapBoostRegressor

X, y = load_diabetes(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42)

model = SnapBoostRegressor(
    num_iterations=100,
    learning_rate=0.1,
    random_state=42,
)
model.fit(X_train, y_train)

print("R²:", model.score(X_test, y_test))
model.evaluate(X_test, y_test)  # prints RMSE

Adaptive training

Version 0.2 adds an opt-in adaptive training path while preserving the original random HNBM algorithm by default:

model = SnapBoostRegressor(
    num_iterations=500,
    learning_rate=0.05,
    selection_strategy="greedy",  # fit the best learner family each round
    line_search=True,              # tune each learner's contribution
    subsample=0.8,                 # stochastic row sampling
    max_features=0.8,              # tree feature sampling
    early_stopping_rounds=30,
    random_state=42,
)
model.fit(
    X_train,
    y_train,
    sample_weight=train_weights,
    eval_set=(X_validation, y_validation),
)

print(model.best_iteration_)
print(model.history_["validation_loss"])

The RFF branch now standardizes its inputs by default and receives a fresh, reproducible random basis each boosting round. Set scale_features=False only when inputs have already been placed on comparable scales.

Optional additive extensions

The classic learner pool remains the default. Additional families and kernels are enabled explicitly:

model = SnapBoostRegressor(
    p_tree=0.7,
    p_linear=0.1,
    kernel_gammas=(0.05, 0.5, 5.0),
    kernel_types=("rbf", "laplacian"),
    objective="pseudo_huber",
    objective_parameter=2.0,
    random_state=42,
)
model.fit(X_train, y_train, candidate_n_jobs=4)

Missing and categorical inputs can be handled outside the estimator with a normal scikit-learn pipeline, keeping SnapBoost's model format unchanged:

from sklearn.pipeline import Pipeline
from snapboost import make_tabular_preprocessor

pipeline = Pipeline([
    ("prepare", make_tabular_preprocessor(categorical_features=(1, 4))),
    ("model", SnapBoostRegressor(random_state=42)),
])

Examples & Results

Interactive Jupyter notebooks in static/ walk through classification, regression, and hyperparameter exploration. Each notebook trains SnapBoost and compares it against XGBoost and LightGBM on the same splits.

Run the notebooks locally:

pip install ".[examples]"
jupyter notebook static/

Classification

On the Breast Cancer dataset (250 boosting rounds), SnapBoost achieves the highest test accuracy and fewest misclassifications among the three boosters:

Confusion matrix for SnapBoost on the held-out test set:

Regression

On the Diabetes dataset (100 boosting rounds), SnapBoost improves R² and RMSE over tree-only baselines:

Predicted vs. actual disease progression on the test set:

SnapBoost fitted curve along BMI (other features held at training medians):

Residual distribution:

Parameter exploration

On a synthetic dataset mixing piecewise-linear and sinusoidal structure, the notebook sweeps p_tree, tree depth ranges, and kernel ridge parameters. A mixed ensemble (p_tree=0.8) outperforms trees-only (p_tree=1.0, RMSE 0.174) and ridge-only (p_tree=0.0, RMSE 0.366):

See Parameter_Exploration.ipynb for the full sweeps and baseline comparison tables.

API Reference

SnapBoostClassifier / SnapBoostRegressor

The recommended entry points (similar to XGBClassifier / XGBRegressor). A concrete HNBM that builds an ensemble from:

• Decision trees with depths sampled uniformly from [min_max_depth, max_max_depth]
• One RFF ridge regressor for smooth global fits

At each iteration, a learner is chosen with probability p_tree for trees (split evenly across depths) and 1 - p_tree for the ridge model.

from snapboost import SnapBoostClassifier, SnapBoostRegressor

clf = SnapBoostClassifier(
    num_iterations=100,
    learning_rate=0.1,
    p_tree=0.8,
    min_max_depth=4,
    max_max_depth=8,
    alpha=1.0,
    gamma=1.0,
    random_state=42,
    verbose=True,
)
clf.fit(X, y)

reg = SnapBoostRegressor(num_iterations=100, random_state=42)
reg.fit(X, y)

Methods

SnapBoost

Legacy class that accepts a mode parameter ("classification" or "regression"). Prefer SnapBoostClassifier or SnapBoostRegressor for new code.

from snapboost import SnapBoost

model = SnapBoost(
    num_iterations=100,
    learning_rate=0.1,
    p_tree=0.8,
    min_max_depth=4,
    max_max_depth=8,
    alpha=1.0,
    gamma=1.0,
    mode="classification",  # or "regression"
    random_state=42,
    verbose=True,
)
model.fit(X, y)

Exact kernel ridge variant

For smaller datasets where an exact RBF kernel is preferable to random Fourier features, task-specific exact-kernel estimators are also available:

from snapboost import (
    SnapBoostKernelRidgeClassifier,
    SnapBoostKernelRidgeRegressor,
)

clf = SnapBoostKernelRidgeClassifier(random_state=42)
reg = SnapBoostKernelRidgeRegressor(random_state=42)

Exact kernel ridge has substantially higher memory and runtime costs than the default RFF learner. The old SnapBoost_KernelRidge name remains available for backward compatibility, but new code should use the task-specific classes.

HNBM

The abstract base class for building custom heterogeneous ensembles. Provided by the hnbm package — subclass or configure base_learners_ and probabilities_ before calling fit:

from sklearn.tree import DecisionTreeRegressor
from hnbm import HNBMClassifier, HNBMRegressor

class MyClassifier(HNBMClassifier):
    def __init__(self, **kwargs):
        super().__init__(**kwargs)
        self.base_learners_ = [DecisionTreeRegressor(max_depth=5)]
        self.probabilities_ = [1.0]

Parameters

Shared (HNBM / SnapBoostClassifier / SnapBoostRegressor)

The legacy SnapBoost class also accepts a mode parameter ("classification" or "regression").

SnapBoost-specific

The adaptive shared parameters are exposed by the recommended SnapBoostClassifier and SnapBoostRegressor classes. Legacy and exact-kernel classes retain their existing constructor surface for compatibility.

After fitting, base_score_ is the optimized constant prediction, learner_weights_ stores per-round contributions, history_ contains training and optional validation loss, and best_iteration_ identifies the round with the lowest validation loss. That ensemble is restored only when early_stopping_rounds triggers; with an eval_set alone best_iteration_ is informational and predictions still use all n_iter_ learners.

Label conventions (classification): accepts any two distinct class labels. Predictions use the original labels, and probability columns follow classes_ order.

Docker

Build and run a container with SnapBoost pre-installed:

docker build -t snapboost .
docker run --rm snapboost

The default command verifies the import:

SnapBoost ready

Development

For normal development against the released HNBM 1.x dependency, install SnapBoost in editable mode and run the complete validation suite:

git clone https://github.com/qiancapital/snapboost.git
cd snapboost
python -m pip install -e ".[test]"
python -m pytest -q
python -m compileall -q snapboost tests

The pytest command must finish with all tests passing. To validate SnapBoost against a local sibling checkout of HNBM, install that checkout first:

python -m pip install -e ../hnbm
python -m pip install -e ".[test]"
python -m pytest -q

Run an individual test module or test while developing with:

python -m pytest -q tests/test_snapboost.py
python -m pytest -q tests/test_rff_learner.py
python -m pytest -q tests/test_snapboost.py::test_classifier_preserves_string_labels

The example notebooks require the separate examples dependencies:

python -m pip install -e ".[examples,test]"
jupyter notebook static/

CI runs the full test suite on every push and pull request, and again before a release distribution is built.

Releases are published to PyPI via GitHub Actions when a GitHub release is created.

References & Citation

If you use this package or the HNBM framework in research, please cite the original SnapBoost paper:

Thomas Parnell, Andreea Anghel, Małgorzata Łazuka, Nikolas Ioannou, Sebastian Kurella, Peshal Agarwal, Nikolaos Papandreou, and Haralampos Pozidis. SnapBoost: A Heterogeneous Boosting Machine. Advances in Neural Information Processing Systems, 33, 2020.

@inproceedings{parnell2020snapboost,
  title     = {{SnapBoost}: A Heterogeneous Boosting Machine},
  author    = {Parnell, Thomas and Anghel, Andreea and {\L}azuka, Ma{\l}gorzata and Ioannou, Nikolas and Kurella, Sebastian and Agarwal, Peshal and Papandreou, Nikolaos and Pozidis, Haralampos},
  booktitle = {Advances in Neural Information Processing Systems},
  volume    = {33},
  pages     = {20872--20883},
  year      = {2020},
  eprint    = {2006.09745},
  doi       = {10.48550/arXiv.2006.09745}
}

Links: arXiv:2006.09745 · NeurIPS proceedings · IBM Research

For the full bibliography, related heterogeneous-boosting literature (KTBoost, DeepBoost, etc.), and notes on how this repo relates to the original IBM Snap ML implementation, see REFERENCES.md. Additional BibTeX entries are in CITATION.bib.

License

MIT License — Copyright (c) 2026 Qian Capital Management LLC (Qian Capital). See LICENSE for full text.