TRexSelector: T-Rex Selector: High-Dimensional Variable Selection & FDR Control

Jasin Machkour, Simon Tien, Daniel P. Palomar, Michael MumaOpenalex Unknown 2022 · DOI · OpenAlex

Abstract

Performs fast variable selection in high-dimensional settings while controlling the false discovery rate (FDR) at a user-defined target level. The package is based on the paper Machkour, Muma, and Palomar (2022) doi:10.48550/arXiv.2110.06048.

Introduction

In modern data science, researchers frequently encounter "large $p$, small $n$" scenarios where the number of potential predictors ($p$) significantly exceeds the number of observations ($n$). In such high-dimensional environments, identifying the true underlying variables is a significant challenge. While standard techniques like the Lasso are effective for promoting sparsity, they often lack robust mechanisms to control the False Discovery Rate (FDR)—the proportion of selected variables that are actually noise.

The TRexSelector R package addresses this critical gap by implementing the T-Rex (Terminating-Random-Experiments) selector. This framework provides a fast, scalable method for variable selection with a built-in guarantee: rigorous control over false positives below a user-defined threshold.

The T-Rex Framework

The core of the package is an intuitive yet mathematically rigorous approach to data augmentation and stability. The T-Rex selector operates through the following stages:

  1. Data Augmentation: The algorithm appends "dummy" predictors—variables consisting of pure noise—to the original predictor matrix ($X$).
  2. Random Experiments: It executes multiple ($K$) random experiments. In each iteration, a forward selection process (such as LARS or Lasso) is performed to track the order in which variables are selected.
  3. Relative Occurrences: The algorithm calculates the frequency with which a real variable is selected before a specific number of dummy variables enter the model.
  4. Calibration and Selection: By comparing the selection behavior of real variables against the known noise baseline (the dummies), the algorithm estimates the False Discovery Proportion (FDP). It then selects the largest set of variables that keeps the estimated FDR below the target level (e.g., 5% or 10%).
Key parameters for the T-Rex selector, including the predictor matrix and dummy variable configuration.

Advanced variants are also available, such as T-Rex+GVS for grouped variable selection and Screen-T-Rex for ultra-fast screening in massive datasets.

Benchmarking with Gauss_data

To demonstrate its capabilities, the package includes a synthetic dataset called Gauss_data. This Gaussian linear model consists of $n = 50$ observations and $p = 100$ variables. The underlying signal is highly sparse, with only 3 variables having non-zero coefficients. This dataset serves as a benchmark for the selector's ability to recover a small signal hidden within a larger field of noise, even when the number of variables is double the number of observations.

Key Insights

  • Dummies as a Calibration Tool: The innovation of T-Rex lies in using dummy variables as a baseline. If a candidate variable cannot consistently outperform random noise in a selection path, it is discarded as a likely false discovery.
  • Stability through Iteration: By defaulting to 20 experiments, the method avoids the instability of single-run selection processes, leading to more reproducible results.
  • User-Defined Risk: Unlike methods that provide a fixed list of variables, T-Rex allows the user to define their tolerance for error, finding the optimal set of predictors within that specific "error budget."
  • Computational Efficiency: The package supports parallel processing via foreach and doParallel, making it viable for large-scale engineering and genomic applications.

Performance and Results

The primary metric for success in the TRexSelector package is the maintenance of the FDR at or below the target level (e.g., tFDR = 0.2). Performance is typically evaluated by comparing the estimated coefficients against the ground truth.

Definitions of regression coefficient vectors used for performance evaluation.

By calculating the True Positive Proportion (TPP) and False Discovery Proportion (FDP) using the estimated vector $\hat{\beta}$ and the true vector $\beta$, users can verify the selector's accuracy. The screen_trex function is particularly noted for its speed in high-dimensional settings where traditional cross-validation is computationally prohibitive.

Limitations and Future Directions

While powerful, the current implementation primarily focuses on Gaussian linear models. Additionally, the requirement to run multiple random experiments introduces a computational overhead compared to a single Lasso path, though this is mitigated by parallelization. Future work may focus on refining how the selector handles extreme collinearity and extending the framework to more complex, non-linear data structures.

Summary updated

Abstract

Performs fast variable selection in high-dimensional settings while controlling the false discovery rate (FDR) at a user-defined target level. The package is based on the paper Machkour, Muma, and Palomar (2022) &lt;<a href="https://doi.org/10.48550/arXiv.2110.06048" target="_top">doi:10.48550/arXiv.2110.06048</a>&gt;.
BibTeX
@inproceedings{machkour2022trexselector,
  author = {Jasin Machkour and Simon Tien and Daniel P. Palomar and Michael Muma},
  title = {{TRexSelector: T-Rex Selector: High-Dimensional Variable Selection \&amp; FDR Control}},
  booktitle = {Openalex Unknown},
  year = {2022},
  doi = {10.32614/cran.package.trexselector},
}
Figure