Scalable XAI approaches for variable interactions interpretability in high dimension

P. Morala Miguélez, E. Gambín Monserrat, I. Úcar Marqués

As AI models grow in complexity, interpretability methods must move beyond assessing the individual importance of input variables and address the nonlinear interactions that often drive model behavior. However, interaction-based explanations, such as SHAP based methods are typically difficult to scale, as their computational cost can increase rapidly with the number of variables considered. This talk presents ongoing work on scalable XAI approaches for interaction interpretability in high-dimensional settings. The central idea is to combine efficient single feature importance techniques with more detailed interaction analysis, applying the latter only to reduced subsets of variables or internal model components selected in the first step. This strategy aims to lower the computational cost of computing interaction importances, by focusing only on the relevant ones and thus ingnoring irrelevant interaction terms and preserving the expressive value of interaction-based explanations.

Palabras clave: Interpretability, ML, AI

GT TABiDa II
2 de septiembre de 2026  12:40
Aula 24

J. A. Cardenas Cardenas, P. Delicado, J. Arjona

S. Díaz-Aranda, J. Aguilar, R. E. Lillo

P. González Barquero, J. Goldsmith, Á. Méndez Civieta, R. E. Lillo