The LASSO-path as a feature importance proxy

V. Moreno, L. Fernández Piana

Feature selection is a fundamental challenge in high-dimensional regression, where identifying the most relevant variables is essential for both interpretability and predictive performance. Although LASSO is widely used for this purpose, its reliance on a single tuning parameter $\lambda$ often limits its ability to recover the true active set. We address this issue by introducing a feature importance score, denoted $I^*$, that leverages the full LASSO path and aggregates information across multiple regularization levels.
We develop two complementary algorithms for computing and applying this score: LASSO.PATH, which efficiently constructs $I^*$, and LASSO.PATH.BISEC, which selects variables by estimating an optimal data-driven threshold. Through simulations and real-data analyses, we demonstrate that our approach improves feature selection accuracy and reduces false discoveries compared with traditional LASSO-based methods.

Keywords: LASSO path, variable selection, feature importance

Data Analysis
September 4, 2026  3:30 PM
Aula 20

J. García-García, M. A. Lubiano

M. Iturrate-Bobes, I. Montes, R. Pérez-Fernández

M. Narbona Sarria

B. Sinova Fernández, I. Cascos Fernández