All You Need for Counterfactual Explainability Is Principled and Reliable Estimate of Aleatoric and Epistemic Uncertainty

TL;DR

The paper argues that transparency research has largely neglected uncertainty quantification, and contends that uncertainty (split into aleatoric and epistemic components) and ante-hoc interpretability are two facets of the same core idea. It proposes that uncertainty provides a principled unifying framework for generating and evaluating counterfactual explanations, enabling more reliable, robust, and human-centered insights. By aligning counterfactual desiderata with uncertainty constraints and adopting path-based explanations, the authors outline a roadmap to build uncertainty-aware, ante-hoc models that can better communicate model limitations and support decision-making in high-stakes settings. The work suggests future directions in uncertainty calibration, second-order uncertainty, and broader explanation types to strengthen the integration of uncertainty into XAI practice, with potential impact in domains like healthcare.

Abstract

This position paper argues that, to its detriment, transparency research overlooks many foundational concepts of artificial intelligence. Here, we focus on uncertainty quantification -- in the context of ante-hoc interpretability and counterfactual explainability -- showing how its adoption could address key challenges in the field. First, we posit that uncertainty and ante-hoc interpretability offer complementary views of the same underlying idea; second, we assert that uncertainty provides a principled unifying framework for counterfactual explainability. Consequently, inherently transparent models can benefit from human-centred explanatory insights -- like counterfactuals -- which are otherwise missing. At a higher level, integrating artificial intelligence fundamentals into transparency research promises to yield more reliable, robust and understandable predictive models.

Figures (5)

• Figure 1: Visualisation of the (\ref{['fig:counterfactual:basic']}) fundamental -- validity, similarity and sparsity -- and (\ref{['fig:counterfactual:extended']}) extended -- plausibility, connectedness, discriminativeness, robustness, stability and actionability -- properties of human-centred counterfactual explanations. The question mark indicates the explained instance and the check mark represents a counterfactual data point.
• Figure 2: Visualisation of counterfactuals in relation to (\ref{['fig:cf:group']}) explanation groups and model selection; (\ref{['fig:cf:cl+rec']}) path-based explainability; and (\ref{['fig:cf:path']}) geometry of the feature space density. The question mark indicates the explained instance, the star symbolises an intermediate step of a counterfactual path, and the check mark represents a counterfactual data point.
• Figure 3: Demonstration of (\ref{['fig:uncertainty:aleatoric']}) aleatoric and (\ref{['fig:uncertainty:epistemic']}) epistemic uncertainty for a two-dimensional toy data set with the model class restricted to linear classifiers hullermeier2021aleatoric.
• Figure 4: Demonstration of how uncertainty quantification depends on the assumed model class for (\ref{['fig:model:linear']}) linear and (\ref{['fig:model:quadratic']}) quadratic classifiers using a two-dimensional toy data set hullermeier2021aleatoric.
• Figure 5: Illustration of uncertainty-driven instance- (green) and path-based (blue) counterfactuals for (\ref{['fig:example-p:linear']}) a standard linear model and (\ref{['fig:example-p:linear+bg']}) one enhanced with a background class perello2016background. In the former case, following the direction perpendicular to the decision boundary is the optimal approach; in the latter case, the counterfactual vector and path are skewed since epistemic uncertainty additionally captures the data manifold shape.