TRexSelector-Python (trexselector)

A Python port of the TRexSelector R package for high-dimensional variable selection with false discovery rate (FDR) control.

Overview

TRexSelector performs fast variable selection in high-dimensional settings while controlling the false discovery rate (FDR) at a user-defined target level. The package implements the Terminating-Random Experiments Selector (T-Rex) as described in Machkour, Muma, and Palomar (2025).

This Python package provides a port of the original R implementation, maintaining the same functionality while providing a more Pythonic interface. The Python port was created by Arnau Vilella ([email protected]).

Installation

Requirements

• Python >= 3.8
• numpy >= 1.20.0
• scipy >= 1.7.0
• scikit-learn >= 1.0.0
• tlars == 0.6.1
• joblib >= 1.0.0
• matplotlib >= 3.4.0
• pandas >= 1.3.0

For manylinux (most Linux distributions), macOS, and Windows platforms, all dependencies including tlars will be installed automatically when installing the package. For other systems, you might need to build the tlars package from source.

pip install trexselector==0.6.4

Usage

import numpy as np
from trexselector import trex, generate_gaussian_data

# Generate some example data
X, y, beta = generate_gaussian_data(n=100, p=20, seed=1234)

# Run the T-Rex selector
res = trex(X=X, y=y)

# Get the selected variables
selected_var = res["selected_var"]
print(f"Selected variables: {selected_var}")

Library Reference

Main Functions

trex(X, y, tFDR=0.2, K=20, max_num_dummies=10, max_T_stop=True, method="trex", ...)

The main function for high-dimensional variable selection with FDR control.

• X: ndarray - Predictor matrix of shape (n, p).
• y: ndarray - Response vector of shape (n,).
• tFDR: float - Target FDR level (between 0 and 1).
• K: int - Number of random experiments.
• max_num_dummies: int - Factor determining maximum number of dummies.
• max_T_stop: bool - If True, maximum number of included dummies is set to ceiling(n/2).
• method: str - Method to use ('trex', 'trex+GVS', 'trex+DA+AR1', 'trex+DA+equi', 'trex+DA+BT', 'trex+DA+NN').
• Returns: dict - Contains selected variables and additional information.

screen_trex(X, y, tFDR=0.2, K=20, max_num_dummies=10, ...)

Screening variant of T-Rex for ultra-high dimensional datasets.

• X, y, tFDR, K: Same as trex().
• q: int - Number of variables to select in each split.
• num_splits: int - Number of splits of the original problem.
• Returns: dict - Contains selected variables and screening information.

random_experiments(X, y, K=20, T_stop=1, num_dummies=None, ...)

Run K random experiments with the T-Rex selector.

• X, y, K: Same as trex().
• T_stop: int - Number of included dummies before stopping.
• num_dummies: int - Number of dummies to append.
• parallel_process: bool - If True, experiments run in parallel.
• Returns: dict - Contains experiment results and statistics.

Helper Functions

add_dummies(X, num_dummies)

Add random dummy variables to the predictor matrix.

• X: ndarray - Predictor matrix.
• num_dummies: int - Number of dummies to append.
• Returns: ndarray - Matrix with appended dummies.

add_dummies_GVS(X, num_dummies, corr_max=0.5)

Add dummy variables with correlation constraints for group variable selection.

• X: ndarray - Predictor matrix.
• num_dummies: int - Number of dummies to append.
• corr_max: float - Maximum allowed correlation between predictors.
• Returns: dict - Contains matrix with dummies and group information.

FDP(beta_hat, beta)

Compute the false discovery proportion.

• beta_hat: ndarray - Estimated coefficient vector.
• beta: ndarray - True coefficient vector.
• Returns: float - False discovery proportion.

TPP(beta_hat, beta)

Compute the true positive proportion.

• beta_hat: ndarray - Estimated coefficient vector.
• beta: ndarray - True coefficient vector.
• Returns: float - True positive proportion.

generate_gaussian_data(n=50, p=100, seed=789)

Generate synthetic Gaussian data for testing.

• n: int - Number of observations.
• p: int - Number of variables.
• seed: int - Random seed.
• Returns: tuple - (X, y, beta) containing predictor matrix, response, and true coefficients.

Advanced Features

The package supports several variants of the T-Rex selector:

• Basic T-Rex: Standard variable selection with FDR control
• T-Rex+GVS: Group variable selection using correlation structure
• T-Rex+DA variants: Dependency-aware variants
  • AR1: Using AR(1) correlation structure
  • Equi: Using equicorrelation structure
  • BT: Using binary tree structure
  • NN: Using nearest neighbor structure

References

• Machkour, J., Muma, M., & Palomar, D. P. (2025). The Terminating-Random Experiments Selector: Fast High-Dimensional Variable Selection with False Discovery Rate Control. Signal Processing, 231, 109894.

License

This package is licensed under the GNU General Public License v3.0 (GPL-3.0).

Acknowledgments

The original R package TRexSelector was created by Jasin Machkour, Simon Tien, Daniel P. Palomar, and Michael Muma. This Python port was developed by Arnau Vilella ([email protected]).