plskit (Python)

Python package for plskit: Partial Least Squares with modern inference, backed by a Rust engine.

What is plskit?

plskit is a Partial Least Squares (PLS) library built around four ideas:

• Modern inference. Confirmatory tests use split-half held-out prediction: split_exact is calibrated by permutation, and split_nb is its faster Fisher-z approximation. Confidence intervals are taken on rotation-invariant quantities (held-out correlation, per-variable leverage, a subsampling z for β), so no Procrustes alignment is needed. The tests are introduced in our NeurIPS 2026 paper.
• Same results in every language. All numerical work runs in one Rust engine. The Python and R wrappers call into it, and the Julia wrapper runs the Python package, so a fixed (version, seed, X, y) gives the same results, within a bit-near tolerance, in every language and on every supported platform.
• Speed. The engine is written in Rust and fits PLS1 with Improved Kernel PLS (Dayal and MacGregor 1997), which makes plskit probably the fastest PLS package on a single core.
• Parallel and RAM-safe. Inference runs efficiently in parallel without blowing up RAM.

Install

pip install plskit

Wheels are built for Linux, macOS, and Windows on recent Python versions (>=3.11). When a wheel is unavailable for your platform, pip falls back to building from the source distribution, which requires a working Rust toolchain.

Quickstart

import numpy as np
import plskit

X = np.random.default_rng(0).normal(size=(200, 20))
y = X[:, :3].sum(axis=1) + np.random.default_rng(1).normal(size=200)

# Fit at fixed K, or let plskit pick K via cross-validation.
model   = plskit.pls1_fit(X, y, k=3, seed=42)
optimal = plskit.pls1_fit(X, y, k="optimal", k_max=10, seed=42)

y_hat = plskit.pls1_predict(model, X_new=X[:5])

# Confirmatory test for any signal; k=1 with split_exact is recommended.
sig = plskit.pls1_confirmatory_test(
    X, y, k=1, test_method="split_exact", seed=42,
)
print(sig.pvalue, sig.statistic)

Public surface

Three model families: PLS1 (single continuous y), sparse PLS1, and PLS3 / PLSSVD (symmetric X/Y covariance analysis) with a sparse variant. PLS2 and multi-block PLS are planned.

• Preprocessing: preprocess.
• PLS1 fit / predict: pls1_fit, pls1_predict; K-selection via pls1_find_k_optimal, pls1_find_k_sequence.
• Sparse PLS1: spls1_fit, spls1_find_keep_optimal, spls1_find_k_optimal, spls1_find_k_sequence (predict with pls1_predict).
• PLS3 / PLSSVD: pls3_fit (alias plssvd_fit), pls3_transform (alias plssvd_transform), sparse spls3_fit (new in 0.6.0).
• Inference: pls1_confirmatory_test, pls3_confirmatory_test, split_nb_gate, pls1_perm_null, pls1_perm_alpha.
• Interpretive: rotate, pls1_rotation_stability.

Results are frozen dataclasses and carry the inputs and seed that produced them.

Errors raised by the Rust engine surface as plskit.PlsKitError (with PlsKitInvalidWeights and PlsKitResamplingDegenerate subclasses) and carry a stable .code for programmatic handling.

This is a thin wrapper

All numerical work runs inside the plskit Rust crate. The Python package converts inputs to f64, calls the engine, and wraps the result. Bug reports and feature requests belong on the monorepo issue tracker.

Documentation

• Python API reference: every function.
• Results reference: every result type.

License

GPL-3.0-or-later.

Citation

@inproceedings{lenartowicz2026cheap,
  title     = {Cheap and Powerful Tests for Supervised Subspaces: Per-Component Inference for {PLS}},
  author    = {Lenartowicz, Pawe{\l} and Plisiecki, Hubert},
  booktitle = {The Fortieth Annual Conference on Neural Information Processing Systems},
  year      = {2026},
  url       = {https://openreview.net/forum?id=xb6CB7d9LO}
}

Paweł Lenartowicz — Freestyler Scientist · GitHub · ORCID