# `bve_select` — BVE — Backward Variable Elimination
_Group_: **Variable selector** · _Registry tolerance_: `1e-06`
## Description
Backward Variable Elimination (§18 Phase 5k)
From the `pls4all.sklearn.BVESelector` docstring:
> Backward Variable Elimination with VIP filter.
> **Registry note** — R `plsVarSel::bve_pls` backward elimination with VIP cut (Mehmood et al.). Default `_bve_select_pls4all` path mirrors the same R call with seed=11, giving bit-exact mask parity. The C++ greedy step-count RMSE backward-elimination kernel is opt-in via `legacy=True`.
### Parameters
| Name | Type | Default | Notes |
|------|------|---------|-------|
| `n_components` | `int` | `2` | Number of latent components extracted (k). |
| `n_steps` | `int` | `10` | Number of elimination passes performed. |
| `min_features` | `int | None` | `None` | Minimum number of variables the selector is allowed to keep (defaults to `n_components`). |
| `n_folds` | `int` | `3` | Number of cross-validation folds used inside the selector. |
## Explanations
### Bibliographic source
Forina, M., Casolino, M. C. & Pizarro Millán, C. (2004). *Iterative predictor weighting (IPW) PLS: a technique for the elimination of useless predictors in regression problems*. Journal of Chemometrics 18(2), 105–112 (§2).
### Mathematical principle
Greedy backward elimination: at each step, evaluate every possible **one-variable removal** by CV-RMSE and drop the feature whose removal hurts least (or helps most). Continue until either a target subset size is reached or removal starts to hurt CV-RMSE materially.
Cost: $O(p^2 \cdot \mathrm{CV})$ — quadratic in the number of features, since each step evaluates $\sim p$ candidate removals. For $p \le 200$ this is tractable; for larger spectra the shaving variant is preferred.
Strength: BVE is essentially exhaustive at each step, so it cannot be tricked by collinearity the way SPA can. Weakness: very expensive on full NIR spectra.
### Implementation
`n4m_feature_selection_bve_select`.
R roxygen note (`methods_extra.R::bve_select`):
> BVE-PLS.
> @param n_components Integer. Number of latent components.
> @param n_steps Method-specific parameter. See the underlying `*_fit()` function for the exact semantics.
> @param min_features Method-specific parameter. See the underlying `*_fit()` function for the exact semantics.
> @param X Numeric matrix of predictors (rows = samples, cols = features).
> @param Y Numeric matrix or vector of responses, with one row per sample.
> @export
### 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**
::::{tab-set}
:class: pls4all-bindings
:::{tab-item} C ABI · libn4m
:sync: c
:class-label: lang-c
```c
/* C ABI — libn4m */
n4m_context_t* ctx = n4m_context_create();
n4m_config_t* cfg = n4m_config_create();
n4m_method_result_t* res = NULL;
n4m_feature_selection_bve_select(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);
```
:::
:::{tab-item} Python · pls4all (raw)
:sync: python-raw
:class-label: lang-python
```python
import pls4all
from pls4all._methods import bve_select_fit
with pls4all.Context() as ctx, pls4all.Config() as cfg:
res = bve_select_fit(ctx, cfg, X, y, n_components=4)
# then: res.matrix("predictions"), res.matrix("coefficients"),
# res.vector("mask"), res.scalar("intercept"), …
```
:::
:::{tab-item} Python · pls4all.sklearn
:sync: python-sklearn
:class-label: lang-python
```python
from pls4all.sklearn import BVESelector
mdl = BVESelector(n_components=2, n_steps=10, min_features=None, n_folds=3)
mdl.fit(X, y)
y_hat = mdl.predict(X_test)
```
:::
:::{tab-item} R · pls4all_method()
:sync: r-dispatcher
:class-label: lang-r
```r
library(pls4all)
# Unified low-level dispatcher (May 2026 R cleanup):
res <- pls4all_method("bve_select", X, y,
n_components = 4L, params = list(n_steps = 35L, min_features = 5L))
# res is a named list with MethodResult arrays/scalars.
# selected_indices / top_k_intervals are 1-based.
```
:::
:::{tab-item} R · pls4all (raw fn)
:sync: r-raw
:class-label: lang-r
```r
library(pls4all)
res <- bve_select(X, Y, n_components, n_steps = 10L, min_features = 5L)
yhat <- pls4all_predict(res, X_test)
```
:::
:::{tab-item} MATLAB · pls4all (MEX)
:sync: matlab-mex
:class-label: lang-matlab
```matlab
res = pls4all.bve_select(X, y, 4);
% see header of bindings/matlab/+pls4all/bve_select.m for full
% parameter surface:
% res = bve_select(X, Y, n_components, n_steps, min_features)
yhat = predict(res, Xtest);
```
:::
:::{tab-item} MATLAB · pls4all (classdef)
:sync: matlab-classdef
:class-label: lang-matlab
_No idiomatic classdef wrapper — invoke `pls4all.fit("bve_select", X, y, …)` directly from the unified MEX factory._
:::
::::
**Registry parity references** 📐
:::{card}
:class-card: external-refs
- 📐 **`ref.r_plsvarsel`** (R · r) — `plsVarSel` 0.10.0 · strict (rmse_rel ≤ 1e-06) — R `plsVarSel::bve_pls` — backward variable elimination with VIP filter.
:::
### Benchmarks
Adaptive wall-clock per cell measured against [`full_matrix.csv`](../benchmarks/overview.md). 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-06`).
Rows tagged with **📐** are the canonical parity references for this method (declared in [`parity_timing.registry`](../benchmarks/methodology.md)). 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.
::::{tab-set}
:class: parity-tabs
:::{tab-item} 1 thread
:sync: threads-1
| Backend | Parity | 200×40 (ms) |
| C++ native · libn4m |
pls4all.cpp.blas+omp | ✓ J 1.00 | 510.0 ms |
| Python · pls4all |
pls4all.python | ✓ J 1.00 | 563.8 ms |
pls4all.sklearn | ⇄ J 0.40 | 42.4 ms🏆 |
| R · pls4all |
pls4all.R | ⇄ J 0.40 | 58.7 ms |
pls4all.R.formula | ⇄ J 0.40 | 66.5 ms |
pls4all.R.mdatools | ⇄ J 0.40 | 152.1 ms |
pls4all.R.pls | ⇄ J 0.40 | 141.0 ms |
| R · external |
📐ref.r_plsvarsel | source | 46.9 ms |
:::
:::{tab-item} 3 threads
:sync: threads-3
| Backend | Parity | 200×40 (ms) |
| C++ native · libn4m |
pls4all.cpp.blas+omp | ✓ J 1.00 | 491.1 ms |
| Python · pls4all |
pls4all.python | ✓ J 1.00 | 491.1 ms |
pls4all.sklearn | ⇄ J 0.40 | 43.0 ms🏆 |
| R · pls4all |
pls4all.R | ⇄ J 0.40 | 61.4 ms |
pls4all.R.formula | ⇄ J 0.40 | 62.6 ms |
pls4all.R.mdatools | ⇄ J 0.40 | 62.1 ms |
pls4all.R.pls | ⇄ J 0.40 | 61.3 ms |
| R · external |
📐ref.r_plsvarsel | source | 47.3 ms |
:::
:::{tab-item} 10 threads
:sync: threads-10
| Backend | Parity | 200×40 (ms) |
| C++ native · libn4m |
pls4all.cpp.blas+omp | ✓ J 1.00 | 507.2 ms |
| Python · pls4all |
pls4all.python | ✓ J 1.00 | 479.8 ms |
pls4all.sklearn | ⇄ J 0.40 | 43.9 ms🏆 |
| R · pls4all |
pls4all.R | ⇄ J 0.40 | 61.3 ms |
pls4all.R.formula | ⇄ J 0.40 | 62.1 ms |
pls4all.R.mdatools | ⇄ J 0.40 | 109.1 ms |
pls4all.R.pls | ⇄ J 0.40 | 60.9 ms |
| R · external |
📐ref.r_plsvarsel | source | 45.8 ms |
:::
::::
---
_See also_: [benchmark overview](../benchmarks/overview.md) · [methods index](index.md) · [interactive dashboard](../landing/dashboard.md)