On Diverse Explanations (Prior to and After Predicting)

E. Carrizosa, R. DE LEONE, M. MAGAGNINI

The so-called Rashomon effect refers to the empirical observation that predictive models with markedly different structural properties can all achieve near-optimal accuracy. Since such models differ, they may yield substantially different explanations.

In the first part of the talk, we study the problem of identifying a set of p predictive models within a given class simultaneously maximizing accuracy and diversity. This leads to a bi-objective optimization problem, closely related to dispersion problems in the Operations Research literature. Several variants, associated with different diversity metrics, are examined.

In the second part, we again aim to identify p predictive models, but shift the notion of diversity from model structure to the explanations generated. Specifically, we seek diversity in the local explanations produced for individual observations, rather than in the properties of the models.

In both settings, the Pareto front is approximated using matheuristics.

Keywords: Predictive models, Explainability, Maximal Dispersion, Biobjective Optimization

GT AMyC III: Mathematical Optimization for Transparent Decision Making
September 4, 2026  3:30 PM
Aula 28

E. Carrizosa, J. C. Castro Gómez, V. Guerrero

M. Alonso-Pena, E. Carrizosa, A. Navas-Orozco

J. Martín-Chávez, E. Carrizosa, C. Molero-Río

C. Molero-Río, C. D'Ambrosio