Joint Optimization of Interpretability, Fairness, and Robustness in Twin SVMs

S. Benítez Peña, S. Novo Díaz, J. García González Redondo

Ensuring interpretability and robustness in machine learning models while preserving fairness is critical for trustworthy decision-making in high-stakes applications. This paper investigates the trade-off between fairness, sparsity, and predictive performance in classification. We propose a robust framework based on Twin SVMs that incorporates sparsity penalties for embedded feature selection together with explicit fairness constraints. Unlike standard approaches, our formulation enables the joint optimization of interpretability and equitable decision-making. To move beyond hard classification toward risk-aware decision-making, we extend the framework to yield calibrated probabilistic predictions. We evaluate ensembles built with different calibration strategies, analyzing their effect on the reliability and stability of probability estimates. Extensive experiments analyze the impact of regularization and fairness levels on both accuracy and the stability of selected features.

Keywords: Algorithmic Fairness, Sparse Learning, Probabilistic Calibration, Interpretable Classification

GT TABiDa I
September 2, 2026  11:20 AM
Aula 24

S. Bianconcini, S. Cagnone, L. Guastadisegni

P. Contró Rodríguez, S. Pineda Sanjuan, C. Ordaz Solera, A. Calviño Martínez

A. Mendez-Civieta, M. C. Aguilera-Morillo, R. E. Lillo