LAMPS

Model-Agnostic Feature Selection via LOCO-Guided Adaptive Minipatch Sampling.

LAMPS selects feature indices using adaptive minipatch sampling. It supports custom prediction functions and grid-search hyperparameter tuning.

Installation

After the first PyPI release:

python -m pip install lamps-fs

The distribution is named lamps-fs; import it as lamps. Requires Python >=3.10 and NumPy >=1.25. NumPy is the only runtime dependency. No R installation or experiment datasets are needed.

Before publication, developers can run python -m pip install . from the repository root.

Quick start

import numpy as np
from lamps import lamps_select

rng = np.random.default_rng(100)
X = rng.normal(size=(80, 20))
y = 4 * X[:, 0] - 3 * X[:, 1] + rng.normal(size=80)

# A small demonstration budget; omit K and max_iter to use the defaults.
selected = lamps_select(X, y, K=300, max_iter=2, show_progress=False)
print(selected)  # Zero-based column indices into X.

X should be a numeric array of shape (n_samples, n_features) and y a one-dimensional numeric response of length n_samples. The default model is NumPy least-squares linear regression without an intercept. Center the data or supply a custom model if you need an intercept.

Tuning and detailed results

Pass a list for n_ratio, m_ratio, fit_func, or delta to search their Cartesian product. LAMPS chooses the run with the lowest final-epoch leave-one-out error and prints its hyperparameters.

selected, results = lamps_select(
    X, y,
    n_ratio=[0.3, 0.4],
    delta=[0.7, 0.8],
    K=300,
    max_iter=2,
    return_complete_info=True,
    show_progress=False,
)

The default return value is an array of selected feature indices. return_complete_info=True returns (selected, results), where results contains iteration-indexed diagnostics. seed controls the sampling randomness; custom models must manage any randomness of their own.

For a custom model, pass a callable fit_func(X_train, y_train, X_predict) that returns one prediction per row of X_predict. Install that model's dependencies separately. A list of callables enables model tuning.

Paper reproduction and license

The GitHub repository contains separate paper experiment scripts and environment instructions. Those scripts, datasets, and vendored comparison methods are not shipped in this package.

LAMPS is distributed under the MIT license, copyright 2026 LAMPS authors.