Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Meta-evaluating stability measures: MAX-Senstivity & AVG-Sensitivity

Miquel Miró-Nicolau, Antoni Jaume-i-Capó, Gabriel Moyà-Alcover

TL;DR

This work tackles the challenge of assessing XAI robustness by introducing a verifiable meta-evaluation framework for stability metrics. It defines two tests, the Perfect Explanation Test ($\mathrm{PET}$) and the Random Output Test ($\mathrm{ROT}$), to benchmark $\mathrm{AVG ext{-}Sensitivity}$ and $\mathrm{MAX ext{-}Sensitivity}$ under controlled conditions. Using a transparent Decision Tree on the TXUXIv3 synthetic dataset and deliberately noisy explanations, the authors show that both metrics pass PET but fail ROT, exposing intrinsic weaknesses due to locality in their definitions. The study provides a practical benchmark to discard unreliable robustness measures and encourages the development of more robust evaluation methodologies for XAI explanations.

Abstract

The use of eXplainable Artificial Intelligence (XAI) systems has introduced a set of challenges that need resolution. The XAI robustness, or stability, has been one of the goals of the community from its beginning. Multiple authors have proposed evaluating this feature using objective evaluation measures. Nonetheless, many questions remain. With this work, we propose a novel approach to meta-evaluate these metrics, i.e. analyze the correctness of the evaluators. We propose two new tests that allowed us to evaluate two different stability measures: AVG-Sensitiviy and MAX-Senstivity. We tested their reliability in the presence of perfect and robust explanations, generated with a Decision Tree; as well as completely random explanations and prediction. The metrics results showed their incapacity of identify as erroneous the random explanations, highlighting their overall unreliability.

Meta-evaluating stability measures: MAX-Senstivity & AVG-Sensitivity

TL;DR

This work tackles the challenge of assessing XAI robustness by introducing a verifiable meta-evaluation framework for stability metrics. It defines two tests, the Perfect Explanation Test () and the Random Output Test (), to benchmark and under controlled conditions. Using a transparent Decision Tree on the TXUXIv3 synthetic dataset and deliberately noisy explanations, the authors show that both metrics pass PET but fail ROT, exposing intrinsic weaknesses due to locality in their definitions. The study provides a practical benchmark to discard unreliable robustness measures and encourages the development of more robust evaluation methodologies for XAI explanations.

Abstract

The use of eXplainable Artificial Intelligence (XAI) systems has introduced a set of challenges that need resolution. The XAI robustness, or stability, has been one of the goals of the community from its beginning. Multiple authors have proposed evaluating this feature using objective evaluation measures. Nonetheless, many questions remain. With this work, we propose a novel approach to meta-evaluate these metrics, i.e. analyze the correctness of the evaluators. We propose two new tests that allowed us to evaluate two different stability measures: AVG-Sensitiviy and MAX-Senstivity. We tested their reliability in the presence of perfect and robust explanations, generated with a Decision Tree; as well as completely random explanations and prediction. The metrics results showed their incapacity of identify as erroneous the random explanations, highlighting their overall unreliability.

Paper Structure

This paper contains 14 sections, 12 equations, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Robustness Measures
  3. Method
  4. Perfect Explanation Test (PET)
  5. Random Output Test (ROT)
  6. Experimental setup
  7. AI Model
  8. Performance measures
  9. Datasets
  10. Experiments
  11. Results and discussion
  12. Experiment 1: Perfection Explanation Test
  13. Experiment 2: Random Output Test
  14. Conclusions

Figures (2)

  • Figure 1: Samples from TXUXIv3 dataset miro2023assessing and their respective explanations from a Decision Tree. We can see the sparse nature of these explanations.
  • Figure 2: Samples from TXUXIv3 dataset miro2023assessing, we can see the different background from DTD dataset cimpoi14describing.