sklears-semi-supervised

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

Overview

sklears-semi-supervised implements semi-supervised learning algorithms that align with scikit-learn’s API, covering label propagation, self-training, and graph-based methods.

Key Features

• Algorithms: LabelPropagation, LabelSpreading, SelfTrainingClassifier, co-training/tri-training, and graph-based methods (harmonic functions, local/global consistency, manifold regularization).
• Graph Support: Efficient knn graph construction, similarity kernels, and Rayon-parallelized graph algorithms (parallel_graph) with SIMD-accelerated distance kernels (simd_distances) for large graphs.
• Pipeline Integration: Works with datasets containing missing labels and plugs into sklears pipelines.
• Monitoring: Built-in tracking for convergence diagnostics and label confidence scores.

Quick Start

use sklears_semi_supervised::LabelSpreading;
use sklears_core::traits::{Fit, Predict};
use scirs2_core::ndarray::{array, Array1};

let x = array![
    [0.0, 1.0],
    [1.0, 0.0],
    [1.0, 1.0],
    [0.5, 0.2],
];
let y = Array1::from(vec![0, 1, -1, -1]); // -1 denotes unlabeled

let model = LabelSpreading::new()
    .kernel("rbf".to_string())
    .gamma(0.5)
    .max_iter(100)
    .tol(1e-3);

let fitted = model.fit(&x.view(), &y.view())?;
let inferred = fitted.predict(&x.view())?;

Status

• Exercised by 356 passing tests in 0.2.0 (Stable).
• Broad coverage of scikit-learn's semi-supervised module (label propagation/spreading, self-training) plus additional graph-based methods (harmonic functions, co/tri-training, manifold regularization) not present upstream.
• Graph algorithms are CPU-parallelized (Rayon) and SIMD-accelerated; no GPU backend is implemented in this crate.
• Additional experiments (semi-supervised regression, curriculum learning) tracked in TODO.md.