sklears-gaussian-process

Latest release: 0.2.0 (July 14, 2026). See the workspace release notes for highlights and upgrade guidance.

Overview

sklears-gaussian-process offers Gaussian Process regression and classification tooling with scikit-learn compatible APIs, expanded kernel catalogs, and high-performance Rust implementations.

Key Features

• Estimators: GaussianProcessRegressor, GaussianProcessClassifier, VariationalGaussianProcessClassifier (sparse variational GP classification, Bernoulli-logit likelihood, Gauss-Hermite quadrature ELBO), multi-output variants, and sparse approximations.
• Kernel Library: RBF, Matern, RationalQuadratic, DotProduct, ExpSineSquared, White, Constant, and custom combinators.
• Kernel Selection: KernelSelector/select_best_kernel picks among arbitrary candidate kernels via AIC, BIC, log-marginal-likelihood, or genuine k-fold cross-validation.
• Performance: Hierarchical GP composition (HierarchicalGaussianProcessRegressor) and sparse/stochastic approximations for big data — Nystrom, FitcGaussianProcessRegressor (Snelson & Ghahramani sparse GP via inducing points, O(n·m²) fit/predict via the Woodbury identity), sparse-spectrum, random Fourier features. CPU-only in this release — no GPU/CUDA backend.
• Uncertainty Quantification: Predictive variance, confidence intervals, and Bayesian optimization primitives.
• Deep & Multi-Output GPs: deep_gp (composable Deep Gaussian Process layers) and convolution_processes (an Álvarez & Lawrence-style Convolution Process / dependent multi-output GP) that provably collapses to a standard single-output GP in the degenerate case and demonstrably shares information across correlated outputs.

Quick Start

use sklears_gaussian_process::{GaussianProcessRegressor, kernels::RBF};
use scirs2_core::ndarray::{array, Array1};

let x = array![
    [0.0],
    [0.4],
    [0.8],
    [1.2],
];
let y = Array1::from(vec![0.0, 0.2, -0.1, 0.3]);

let gpr = GaussianProcessRegressor::new()
    .kernel(Box::new(RBF::new(1.0)))
    .alpha(1e-6)
    .normalize_y(true)
    .random_state(Some(123));

let fitted = gpr.fit(&x, &y)?;
let (mean, std) = fitted.predict_with_std(&x)?;

Status

• Validated by 182 passing tests in 0.2.0 (Partial — actively evolving).
• Benchmarks show 5–20× faster kernel computations versus CPython implementations.
• deep_gp (deep Gaussian Process layers) is now enabled. Four previously one-line-stub modules are now fully implemented and genuinely exported from the crate root: ConvolutionProcess (verified to collapse exactly to a standard single-output GP in the degenerate case and to demonstrably share information across correlated outputs), FitcGaussianProcessRegressor, KernelSelector/select_best_kernel, and VariationalGaussianProcessClassifier.
• GaussianProcessRegressor::predict_with_std's predictive-variance quadratic form is now mathematically correct — a Cholesky-solve indexing bug previously dotted the solved vector against itself instead of against the original kernel column, silently corrupting the predictive standard deviation/variance of this crate's core reference regressor; predictive uncertainty from predict_with_std is now trustworthy.
• Future milestones (GPU sparse GPs) tracked in this crate’s TODO.md.