Angel D. Reyero Lobo

Angel D. Reyero Lobo

PhD candidate in Machine Learning at Institut de Mathématiques de Toulouse/ MIND, Inria Paris-Saclay

Publications

Conference Papers

Aggregate Models, Not Explanations: Improving Feature Importance Estimation

Published in International Conference on Machine Learning (ICML 2026), 2026

We show that model-level ensembling improves the stability and accuracy of feature-importance estimates for expressive machine-learning models, providing theoretical guarantees and empirical validation on benchmarks and UK Biobank proteomic data.

Recommended citation: Joseph Paillard, Angel Reyero Lobo, Denis A. Engemann, Bertrand Thirion. Aggregate Models, Not Explanations: Improving Feature Importance Estimation. In Proceedings of the International Conference on Machine Learning (ICML), 2026.
Download Paper

Semi-knockoffs: a model-agnostic conditional independence testing method with finite-sample guarantees

Published in International Conference on Machine Learning (ICML 2026), 2026

We propose a conditional independence testing procedure with valid type-I error and FDR control that accommodates any pretrained model and requires no train–test split.

Recommended citation: Angel Reyero-Lobo, Bertrand Thirion, and Pierre Neuvial. Semi-knockoffs: a model-agnostic conditional independence testing method with finite-sample guarantees. In Proceedings of the International Conference on Machine Learning (ICML), 2026.
Download Paper

A principled approach for comparing Variable Importance

Published in Preprint, 2025

The paper introduces an axiomatic framework and a principled approach to create and evaluate variable importance measures (VIMs), ensuring they avoid spurious correlations and enable fair comparisons. It also offers examples to help practitioners choose and estimate suitable VIMs for their goals and data.

Recommended citation: Reyero-Lobo, A., Neuvial, P. , & Thirion, B. (2025). A principled approach for comparing Variable Importance.
Download Paper

When Knockoffs fail: diagnosing and fixing non-exchangeability of Knockoffs

Published in Preprint, 2025

This work identifies and addresses failures of the exchangeability assumption in knockoff-based inference by introducing a diagnostic test and proposing a robust alternative construction method.

Recommended citation: Blain, A., Reyero Lobo, A., Linhart, J., Thirion, B., & Neuvial, P. (2025). When Knockoffs fail: Diagnosing and fixing non-exchangeability of Knockoffs.
Download Paper

Measuring Variable Importance in Heterogeneous Treatment Effects with Confidence

Published in ICML, 2025

PermuCATE is a new method for assessing variable importance in estimating heterogeneous treatment effects. It improves on existing techniques by reducing variance, making it more reliable—especially in low-data settings like biomedical applications.

Recommended citation: Paillard, J., Reyero Lobo, A., Kolodyazhniy, V., Thirion, B., & Engemann, D. A. (2025). Measuring variable importance in heterogeneous treatment effects with confidence.
Download Paper

Analyse statistique d’hill climbers à voisinage large pour le problème d’ordonnancement linéaire

Published in ROADEF24, 2024

Use of linear programming and insertion algorithms to define neighborhoods and obtain local optimums. Ultimately, perturb and modify neighborhoods to enhance solutions.

Recommended citation: Angel David Reyero Lobo, Nicolas Dupin. Analyse statistique d’hill climbers à voisinage large pour le problème d’ordonnancement linéaire. ROADEF 2024, Mar 2024, Amiens, France. ffhal-04450707
Download Paper