vissa_select — VISSA — Variable Iterative Space-Shrinkage

Group: Variable selector · Registry tolerance: 1e-06

Description

VISSA-PLS — Variable Iterative Space Shrinkage (§49)

Registry note — Python auswahl.VISSA 0.9.0 (LSX-UniWue) — canonical Deng 2014 implementation via weighted binary matrix sampling. Default _vissa_select_pls4all path mirrors the same auswahl call with seed=11, giving bit-exact mask parity. The C++ splitmix64 VISSA kernel is opt-in via legacy=True.

Parameters

Name	Type	Default	Notes
n_components	int	2	current public binding signature
n_iterations	int	10	current public binding signature
n_submodels	int	60	current public binding signature
ratio_kept	float	0.1	current public binding signature
threshold	float	0.5	current public binding signature
floor_probability	float	0.05	current public binding signature
n_folds	int	3	current public binding signature
seed	int	0	current public binding signature

Explanations

Bibliographic source

Deng, B. C., Yun, Y. H., Liang, Y. Z. & Yi, L. Z. (2015). A new strategy to prevent over-fitting in partial least squares models based on model population analysis. Analytica Chimica Acta 880, 32–41. DOI 10.1016/j.aca.2015.04.045.

Mathematical principle

VISSA evaluates a population of random subsets of the same size, refines the population by selecting the best by CV-RMSE, and iteratively shrinks the search space toward features that survive in many high-performing subsets. Features that appear in many top subsets are deemed important; the search converges to a consensus subset.

Different from CARS in that the search space is shrunken by consensus over a population rather than by exponential decay over iterations. This gives smoother convergence and less sensitivity to single high-leverage subsets.

Appropriate uses

Iteratively concentrating weighted sampling on wavelengths associated with better PLS subsets.

Limits and validation

Sampling probabilities are path-dependent and stochastic; reported subsets need independent assessment.

Implementation

n4m_feature_selection_vissa_select.

Sources and provenance

Current implementation: cpp/src/core/vissa_selection.cpp.

API and bindings

C ABI (ABI 2): n4m_feature_selection_vissa_select. Use the linked public header for the exact signature, configuration, and result handles.

Python (verified public re-export):

from n4m.feature_selection.wrapper import VISSA

Source signature: VISSA(*, n_components: int = 2, n_iterations: int = 10, n_submodels: int = 60, ratio_kept: float = 0.1, threshold: float = 0.5, floor_probability: float = 0.05, n_folds: int = 3, seed: int = 0) (n4m/_impl/selection.py).

R (source-verified): vissa_select(X, Y, n_components, n_iterations = 20L, n_submodels = 100L, ratio_kept = 0.1, threshold = 0.5, floor_probability = 0.01, seed = 0L).

The source signature has additional required inputs, so no example call is fabricated.

MATLAB / Octave (source-verified): vissa_select(X, Y, n_components, n_iterations, n_submodels, ... ratio_kept, threshold, floor_probability, seed).

The source signature has additional required inputs, so no example call is fabricated.

Registry parity references 📐

• 📐 ref.python_auswahl (python · python) — auswahl 0.9.0 · strict (rmse_rel ≤ 1e-06) — Python auswahl.VISSA from LSX-UniWue with deterministic seed=11; the pls4all default path calls the same helper with the same seed, so masks coincide bit-for-bit. The C++ splitmix64 VISSA kernel is opt-in via legacy=True.

Benchmarks

Archived measurement identity. Backend labels in this table are the raw IDs recorded when the benchmark ran (including historical pls4all.* IDs). They preserve measurement provenance and do not describe a current public Python, R, or MATLAB binding; use the source-verified API and bindings section above for current entry points.

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 = archived binding-harness result 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). C++ and external rows show reference parity; archived language-harness rows show binding parity against the C++ backend. Hover the icon for role and tolerance band.

1 thread

Backend	Parity	80×25 (ms)
Python · archived pls4all benchmark
pls4all.sklearn	✓ J 0.60	10.8 ms🏆
R · archived pls4all benchmark
pls4all.R	✓ J 0.60	11.2 ms
pls4all.R.formula	✓ J 0.60	11.1 ms
pls4all.R.mdatools	✓ J 0.60	49.1 ms
pls4all.R.pls	✓ J 0.60	21.4 ms

3 threads

Backend	Parity	80×25 (ms)
Python · archived pls4all benchmark
pls4all.sklearn	✓ J 0.60	13.0 ms
R · archived pls4all benchmark
pls4all.R	✓ J 0.60	13.9 ms
pls4all.R.formula	✓ J 0.60	14.6 ms
pls4all.R.mdatools	✓ J 0.60	11.4 ms🏆
pls4all.R.pls	✓ J 0.60	11.4 ms

10 threads

Backend	Parity	80×25 (ms)
Python · archived pls4all benchmark
pls4all.sklearn	✓ J 0.60	8.85 ms🏆
R · archived pls4all benchmark
pls4all.R	✓ J 0.60	11.8 ms
pls4all.R.formula	✓ J 0.60	12.5 ms
pls4all.R.mdatools	✓ J 0.60	13.1 ms
pls4all.R.pls	✓ J 0.60	11.2 ms

---

See also: benchmark overview · methods index · interactive dashboard