Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Do Metrics for Counterfactual Explanations Align with User Perception?

Felix Liedeker, Basil Ell, Philipp Cimiano, Christoph Düsing

Abstract

Explainability is widely regarded as essential for trustworthy artificial intelligence systems. However, the metrics commonly used to evaluate counterfactual explanations are algorithmic evaluation metrics that are rarely validated against human judgments of explanation quality. This raises the question of whether such metrics meaningfully reflect user perceptions. We address this question through an empirical study that directly compares algorithmic evaluation metrics with human judgments across three datasets. Participants rated counterfactual explanations along multiple dimensions of perceived quality, which we relate to a comprehensive set of standard counterfactual metrics. We analyze both individual relationships and the extent to which combinations of metrics can predict human assessments. Our results show that correlations between algorithmic metrics and human ratings are generally weak and strongly dataset-dependent. Moreover, increasing the number of metrics used in predictive models does not lead to reliable improvements, indicating structural limitations in how current metrics capture criteria relevant for humans. Overall, our findings suggest that widely used counterfactual evaluation metrics fail to reflect key aspects of explanation quality as perceived by users, underscoring the need for more human-centered approaches to evaluating explainable artificial intelligence.

Do Metrics for Counterfactual Explanations Align with User Perception?

Abstract

Explainability is widely regarded as essential for trustworthy artificial intelligence systems. However, the metrics commonly used to evaluate counterfactual explanations are algorithmic evaluation metrics that are rarely validated against human judgments of explanation quality. This raises the question of whether such metrics meaningfully reflect user perceptions. We address this question through an empirical study that directly compares algorithmic evaluation metrics with human judgments across three datasets. Participants rated counterfactual explanations along multiple dimensions of perceived quality, which we relate to a comprehensive set of standard counterfactual metrics. We analyze both individual relationships and the extent to which combinations of metrics can predict human assessments. Our results show that correlations between algorithmic metrics and human ratings are generally weak and strongly dataset-dependent. Moreover, increasing the number of metrics used in predictive models does not lead to reliable improvements, indicating structural limitations in how current metrics capture criteria relevant for humans. Overall, our findings suggest that widely used counterfactual evaluation metrics fail to reflect key aspects of explanation quality as perceived by users, underscoring the need for more human-centered approaches to evaluating explainable artificial intelligence.
Paper Structure (16 sections, 8 equations, 3 figures, 2 tables)

This paper contains 16 sections, 8 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Data Acquisition and Study Design
  4. Datasets and Counterfactual Generation
  5. Sampling of Counterfactual Explanations
  6. User Study Procedure
  7. Power Analysis
  8. Rating Aggregation and Reliability
  9. Automated Metrics
  10. Results
  11. Metric--Rating Correlations
  12. Predictive Modeling
  13. Discussion
  14. Conclusion
  16. ...and 1 more sections

Figures (3)

  • Figure 1: Metric--rating correlations including the CQS for (a) MUS, (b) OBE, (c) HRT datasets. Significant correlations ($p<0.05$) are marked with an asterisk.
  • Figure 2: Histograms of $R^2$ for (a) linear regression and (b) the best performing non-linear model, RF, on the HRT dataset predicting user satisfaction rating.
  • Figure 3: $R^2$ by number of included metrics for (a) linear regression and (b) RF, the best performing non-linear model, on the HRT dataset predicting user satisfaction rating. Lines show the best and mean $R^2$ across all metric combinations at each complexity level, with shaded bands indicating $\pm 1$ SD.