group_sparse_pls — Group-sparse PLS (Liquet 2016)

Group: Sparse · Registry tolerance: 0.05

Description

Group sparse PLS (§7)

From the pls4all.sklearn.GroupSparsePLSRegression docstring:

Group-sparse PLS — L1 across pre-declared feature groups.

Registry note — R sgPLS::gPLS (Liquet et al. 2016, regression mode, scale=TRUE). pls4all’s default kernel is a deterministic NumPy port of this algorithm (shared with _GroupSparseNumpyReference) and agrees with the R reference to ~1e-14. The original C++ soft-threshold-on-weights kernel is opt-in via legacy=True.

Parameters

Name

Type

Default

Notes

n_components

int

2

Number of latent components extracted (k).

group_assignment

—

None

Integer array assigning each feature to a group (length n_features).

group_lambda

float

0.05

L1 penalty applied at the group level (group-sparse PLS).

Explanations

Bibliographic source

Liquet, B., de Micheaux, P. L., Hejblum, B. P. & Thiébaut, R. (2016). Group and sparse group partial least squares approaches applied in genomics context. Bioinformatics 32(1), 35–42.

Mathematical principle

When features partition into known groups — gene pathways, spectroscopic bands, biological assays — group-sparse PLS forces entire groups in or out together via a group-lasso penalty: \(\mathcal{P}(\mathbf{w}) = \sum_g \sqrt{|g|}\,\|\mathbf{w}_g\|_2\), where \(\mathbf{w}_g\) is the sub-vector of weights belonging to group \(g\) and \(|g|\) is its size. The \(\ell_2\) norm inside the sum is non-differentiable at zero, which produces group-level sparsity (an entire \(\mathbf{w}_g\) is either zero or non-zero).

Compared to plain sparse PLS, this gives a much more interpretable model when groups have biological meaning and avoids the situation where one or two members of a co-regulated cluster get selected while the rest don’t.

Required input: a group_assignment vector mapping each feature to a group id.

Implementation

n4m_group_sparse_pls_fit. Reference: CRAN sgPLS 1.8.1.

MATLAB header (bindings/matlab/+pls4all/group_sparse_pls.m):

pls4all.group_sparse_pls  Group-sparse PLS (group L1 over feature groups).

Usage

Every pls4all binding tab dispatches into the same C kernel; the external libraries listed at the bottom of the page are the parity references registered in benchmarks.parity_timing.registry. Switch tabs to read the same fit in your language. The R package now ships drop-in-compatible facades for the CRAN pls package (plsr, pcr, mvr) and for the mdatools::pls(x, y, ...) matrix idiom — those tabs appear only on the methods that have a meaningful equivalence.

pls4all bindings

/* C ABI — libn4m */
n4m_context_t* ctx = n4m_context_create();
n4m_config_t*  cfg = n4m_config_create();
n4m_method_result_t* res = NULL;
n4m_group_sparse_pls_fit(ctx, cfg, &x_view, &y_view, /* hyperparams */, &res);
/* … read coefficients / mask / scores via */
/* n4m_method_result_get_double_matrix / vector / scalar … */
n4m_method_result_destroy(res);
n4m_config_destroy(cfg);
n4m_context_destroy(ctx);

import pls4all
from pls4all._methods import group_sparse_pls_fit
with pls4all.Context() as ctx, pls4all.Config() as cfg:
    res = group_sparse_pls_fit(ctx, cfg, X, y, n_components=4, group_assignment=groups)
# then: res.matrix("predictions"), res.matrix("coefficients"),
# res.vector("mask"), res.scalar("intercept"), …

from pls4all.sklearn import GroupSparsePLSRegression
mdl = GroupSparsePLSRegression(n_components=2, group_assignment=None, group_lambda=0.05)
mdl.fit(X, y)
y_hat = mdl.predict(X_test)

library(pls4all)
# Unified low-level dispatcher (May 2026 R cleanup):
res <- pls4all_method("group_sparse_pls", X, y,
                      n_components = 4L, params = list(group_lambda = 0.1))
# res is a named list with MethodResult arrays/scalars.
# selected_indices / top_k_intervals are 1-based.

res = pls4all.group_sparse_pls(X, y, 4);
% see header of bindings/matlab/+pls4all/group_sparse_pls.m for full
% parameter surface:
%   res = group_sparse_pls(X, Y, n_components, group_assignment, group_lambda)
yhat = predict(res, Xtest);

No idiomatic classdef wrapper — invoke pls4all.fit("group_sparse_pls", X, y, …) directly from the unified MEX factory.

Registry parity references 📐

  • 📐 ref.python_numpy (python · python) — numpy in-tree · relaxed (rmse_rel ≤ 5e-02) — In-tree NumPy port of Liquet et al. 2016 group sparse PLS (R sgPLS::gPLS, regression mode, scale=TRUE). pls4all’s default wrapper calls the same function, so the parity gate is bit-for-bit (max_abs < 1e-6). R sgPLS::gPLS is the published algorithmic counterpart and also matches to double-precision; the legacy C++ kernel (SIMPLS + soft-threshold-on-weights) is opt-in via legacy=True.

  • 📐 ref.r_sgpls (R · r) — sgPLS 1.8.1 · relaxed (rmse_rel ≤ 5e-02) — R sgPLS::gPLS(X, Y, ncomp, ind.block.x, keepX=length(bnd)) (regression, scale=TRUE). The pls4all default kernel is a deterministic NumPy port of this algorithm and agrees to 1e-14 against this reference.

Benchmarks

Adaptive wall-clock per cell measured against full_matrix.csv. Only backends that implement this method are listed; libraries without the method are omitted.

Verdict  ·  ✓ ref / ≈ ref / ~ shape mark a reference-gate pass at strict / relaxed / qualitative tolerance  ·  ✓ bind = pls4all binding agrees with the C++ baseline  ·  ⇄ cross-check = documented by-design selector/RNG/model, noncanonical API/facade convention, or secondary oracle  ·  ✗ divergent  ·  ⚠ error  ·  — not run. The fastest backend per column is marked 🏆.

Reference gate: strict — numeric equivalence (rmse_rel_tol ≤ 1e-08).

Rows tagged with 📐 are the canonical parity references for this method (declared in parity_timing.registry). C++ and external rows show reference parity; pls4all language bindings show binding parity against the C++ backend. Hover the icon for role and tolerance band.

BackendParity200×50 (ms)
C++ native · libn4m
pls4all.cpp.blas+omp✓ ref8.36 ms
Python · pls4all
pls4all.python✓ bind7.86 ms
pls4all.sklearn⇄ +1e-013.79 ms
R · pls4all
pls4all.R⇄ +1e-019.70 ms
pls4all.R.formula⇄ +1e-0111.0 ms
pls4all.R.mdatools⇄ +1e-0110.2 ms
pls4all.R.pls⇄ +1e-0111.0 ms
Python · external
📐ref.python_numpysource2.98 ms🏆
R · external
📐ref.r_sgpls⇄ +1e-1533.7 ms
BackendParity200×50 (ms)
C++ native · libn4m
pls4all.cpp.blas+omp✓ ref2.62 ms
Python · pls4all
pls4all.python✓ bind2.76 ms
pls4all.sklearn⇄ +1e-012.28 ms🏆
R · pls4all
pls4all.R⇄ +1e-018.78 ms
pls4all.R.formula⇄ +1e-019.80 ms
pls4all.R.mdatools⇄ +1e-0111.9 ms
pls4all.R.pls⇄ +1e-0110.5 ms
Python · external
📐ref.python_numpysource3.11 ms
R · external
📐ref.r_sgpls⇄ +1e-1531.2 ms
BackendParity200×50 (ms)
C++ native · libn4m
pls4all.cpp.blas+omp✓ ref2.91 ms
Python · pls4all
pls4all.python✓ bind2.69 ms
pls4all.sklearn⇄ +1e-012.53 ms🏆
R · pls4all
pls4all.R⇄ +1e-017.45 ms
pls4all.R.formula⇄ +1e-019.19 ms
pls4all.R.mdatools⇄ +1e-0111.3 ms
pls4all.R.pls⇄ +1e-018.50 ms
Python · external
📐ref.python_numpysource2.66 ms
R · external
📐ref.r_sgpls⇄ +1e-1526.7 ms

See also: benchmark overview · methods index · interactive dashboard