CONFIDERAI: CONFormal Interpretable-by-Design score function for Explainable and Reliable Artificial Intelligence - Extended Abstract

Aug 3, 2026

Upload by Asher

Trustworthiness in artificial intelligence is framed through two consistent pillars: explainability and conformity. This extended abstract presents a framework that merges both concepts by building conformal rule-based predictions. It introduces a new score function for rule-based models, combining rule relevance with the geometrical position of points relative to rule classification boundaries. The approach yields a restricted conformal safety set, providing probabilistic guarantees on performance within regions where the underlying XAI model remains reliable.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages4

Reading Time10 min

AccessFree to download

Introduction

CONFIDERAI

Conformal Safety Set

Novel Score Function

Preliminary Results

Acknowledgements

References

POLITECNICO DI TORINO Repository ISTITUZIONALE CONFIDERAI: CONFormal Interpretable-by-Design score function for Explainable and Reliable Artificial Intelligence Original CONFIDERAI: CONFormal Interpretable-by-Design score function for Explainable and Reliable Artificial Intelligence / Narteni, Sara; Carlevaro, Alberto; Muselli, Marco; Dabbene, Fabrizio; Mongelli, Maurizio. -ELETTRONICO. -204:(2023), pp. 1-3. (Intervento presentato al convegno The 12th Symposium on Conformal and Probabilistic Prediction with Applications (COPA 2023) tenutosi a Limassol (CY) nel 13-15 September 2023) . Availability: This version is available at: 11583/2982243 since: 2023-09-22T07:04:03Z Publisher: Proceedings of Machine Learning Research Published DOI: Terms of use: This article is made available under terms and conditions as specified in the corresponding bibliographic description in the repository Publisher copyright (Article begins on next page) 18 September 2024 CONFIDERAI: CONFormal Interpretable-by-Design score function for Explainable and Reliable Artificial Intelligence Sara Narteni 1 ,† Alberto Carlevaro 1 ,† Fabrizio Dabbene 1 Marco Muselli 1 ,2 Maurizio Mongelli 1 1 CNR-IEIIT, 10129, Turin, Italy sara.narteni@ieiit.cnr.it alberto.carlevaro@ieiit.cnr.it fabrizio.dabbene@ieiit.cnr.it marco.muselli@ieiit.cnr.it maurizio.mongelli@ieiit.cnr.it 2 Rulex Innovation Labs, Rulex Inc. , 16122 Genoa, Italy † S. Narteni and A. Carlevaro contributed equally to the development of the article. (Corresponding authors: S.Narteni, A. Carlevaro.) Editor: Harris Papadopoulos, Khuong An Nguyen, Henrik Boström and Lars Carlsson Abstract The concept of trustworthiness has been declined in different ways in the field of artificial intelligence, but all its definitions agree on two main pillars: explainability and conformity. In this extended abstract, our aim is to give an idea on how to merge these concepts, by defining a new framework for conformal rule-based predictions. In particular, we introduce a new score function for rule-based models, that leverages on rule relevance and geometrical position of points from rule classification boundaries. Keywords: XAI, conformal safety sets, novel score function, conformal prediction. 1. Introduction Literature around combination of eXplainable AI (XAI) and conformal prediction (CP) has recently gained popularity but it still remains little investigated in research. Some relevant approaches proposed so far investigated conformal prediction for XAI models (Bhattacharyya, 2011 ; Johansson et al. , 2014 , 2018 , 2022), but to the best of our knowledge, no previous study of this type addressed score functions and quantile, tailored for rule-based models. For this reason, we propose CONFIDERAI, an innovative approach, based on a new score function, to build conformal prediction of rule-based models. The rationale behind the approach is the combination of the global properties of decision rules (i.e., their covering and error) and the geometrical position of the points inside rule boundaries. The resulting prediction set leads to a restricted conformal safety set, i.e., the set of points for which the underlying XAI model performs with probabilistic guarantees. 2. CONFIDERAI Conformal Safety Set. CSS allows to insert CP in a more safety-based context. For for any input feature x ∈ X and any label y ∈ Y, given a prediction set at level of confidence 1 − ε , ε ∈ (0 , 1) , C (x) = {y | s (x, y) ≤ sε} ∈ 2Y , (1) © 2023 S. Narteni, A. Carlevaro, F. Dabbene, M. Muselli & M. Mongelli. Narteni Carlevaro Dabbene Muselli Mongelli where sε is the 1 − ε quantile of the score values computed on a calibration set, CSS is defined as a subset of the input feature space in which probabilistic safety guarantees can be provided to the machine learning (ML) model: Sε = {x | Pr{y ∈ C (x)} ≥ 1 − ε, ∀y ∈ Y} = {x | s (x, y) ≤ sε , ∀y ∈ Y} . (2) Novel Score Function. An inn

FAQs

What two pillars define trustworthiness in artificial intelligence in this work?

The abstract states that all definitions agree on two main pillars: explainability and conformity.

How does CONFIDERAI combine explainability and conformity for rule-based predictions?

It defines a new framework for conformal rule-based predictions, introducing a score function that uses rule relevance and the geometrical proximity of points to rule boundaries.

What is a Conformal Safety Set (CSS) and what does it guarantee?

CSS is a subset of the feature space where probabilistic safety guarantees can be provided, expressed through confidence-controlled prediction sets with coverage at level 1 − ε.

Aug 3, 2026

Upload by Asher