Uncertainty Gating for Cost-Aware Explainable Artificial Intelligence

Abstract

Post-hoc explanation methods are widely used to interpret black-box predictions, but their generation is often computationally expensive and their reliability is not guaranteed. We propose epistemic uncertainty as a low-cost proxy for explanation reliability: high epistemic uncertainty identifies regions where the decision boundary is poorly defined and where explanations become unstable and unfaithful. This insight enables two complementary use cases: `improving worst-case explanations' (routing samples to cheap or expensive XAI methods based on expected explanation reliability), and `recalling high-quality explanations' (deferring explanation generation for uncertain samples under constrained budget). Across four tabular datasets, five diverse architectures, and four XAI methods, we observe a strong negative correlation between epistemic uncertainty and explanation stability. Further analysis shows that epistemic uncertainty distinguishes not only stable from unstable explanations, but also faithful from unfaithful ones. Experiments on image classification confirm that our findings generalize beyond tabular data.

Figures (8)

• Figure 1: Uncertainty-guided XAI resource allocation via epistemic gating. Epistemic uncertainty is obtained from the model's native estimator or a lightweight surrogate. Under constrained computation budget, Use Case 1 routes samples to low- or high-cost XAI methods based on expected reliability; Use Case 2 defers high-uncertainty samples, retaining only reliable explanations.
• Figure 2: Correlation heatmaps of XAI stability and epistemic uncertainty (XEC) under natural (left) and adversarial (right) perturbations. Each cell reports Spearman correlation between epistemic growth (EG) and explanation degradation (XD) across perturbation strengths for a given dataset, model and XAI method. Blue and gray regions ($\text{XEC} < -0.6$) correspond to strong negative association, where UQ serves as a reliable proxy for XAI instability.
• Figure 3: Stratified SHAP stability (Kendall’s $\tau$) across epistemic uncertainty bins under Gaussian noise ($\sigma$) for low (blue), medium (green), and high (red) epistemic strata. Violin plots show the distribution of stability values. Dots indicate means, horizontal bars the median, and vertical bars the standard deviation.
• Figure 4: Prediction shift (MSE in log-odds space) after removing top-$k$ SHAP features by low- (blue), high- (red), and random-epistemic (gray) groups. Low-epistemic samples show larger shifts, indicating more faithful explanations.
• Figure 5: Epistemic uncertainty (left axis, blue) and SSIM stability (right axis, orange) means and standard deviations versus noise level (bottom axis). Axes are color-coded. As noise increases, uncertainty rises while SSIM decreases monotonically (XEC $\rho = -1.0$ for both IG and SG).