Uncertainty Gating for Cost-Aware Explainable Artificial Intelligence

Georgii Mikriukov; Grégoire Montavon; Marina M. -C. Höhne

cs.AIMar 31, 2026

Uncertainty Gating for Cost-Aware Explainable Artificial Intelligence

Authors: Georgii Mikriukov, Grégoire Montavon, Marina M. -C. Höhne

Organizations: Leibniz Institute for Agricultural Engineering and Bioeconomy (ATB), Potsdam, Germany · BIFOLD – Berlin Institute for the Foundations of Learning and Data, Berlin, Germany · Charité – Universitätsmedizin Berlin, Berlin, Germany · University of Potsdam – Department of Computational Science, Potsdam, Germany

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 & tables

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.

Type	Dataset	Features	Classes	Train	Val	Test
Tabular	Wine Quality	11	2	4547	975	975
	Dry Bean	16	7	9527	2042	2042
	Rice	7	2	2667	571	572
	Ecoli	7	8	235	50	51
Image	PlantVillage	128 $\times$ 128 $\times$ 3	3	3936	843	847

Table 1: Dataset statistics and train/validation/test splits used in experiments.

Dataset	Model	F1 (Non-UQ)	F1 (UQ)	Epistemic	$\text{CV}_{\text{epi}}$
Wine	LR	0.736	0.733	0.0004	1.40
	RF	0.810	0.810	0.083	0.50
	MLP	0.764	0.773	0.023	0.71
	LGBM	0.808	0.806	–	–
	CB	0.780	0.780	–	–
Bean	LR	0.921	0.923	0.0001	3.71

Table 2: Predictive performance and uncertainty statistics. F1 (Non-UQ) corresponds to deterministic models without uncertainty estimation. Epistemic: mean test-set epistemic uncertainty;

\text{CV}_{\text{epi}}

: coefficient of variation measuring uncertainty dispersion. GBDTs epistemic is not evaluated (–).

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.

	Wine		Bean		Rice
$\nu$	Prec.	Rec.	Prec.	Rec.	Prec.	Rec.
90%	0.932	0.144	1.000	0.123	1.000	0.114
70%	0.788	0.367	1.000	0.370	1.000	0.342
50%	0.735	0.570	0.996	0.614	1.000	0.570
30%	0.687	0.746	0.989	0.854	0.998	0.797
10%	0.648	0.904	0.890	0.987	0.937	0.962

Table 3: Precision and recall for detecting stable explanations (

\tau\geq 0.7

) at different deferral rates

\nu

under pooled noise conditions.

		RF + TreeSHAP		MLP + LIME
Dataset	$\nu$	Stability ( $\tau$ )	$q$	Stability ( $\tau$ )	$q$
Wine	70%	$0.797\pm 0.121$	0.30	$0.722\pm 0.106$	0.31
	50%	$0.777\pm 0.130$	0.50	$0.725\pm 0.108$	0.51
	30%	$0.758\pm 0.141$	0.70	$0.729\pm 0.108$	0.71
	0%	$0.740\pm 0.149$	1.00	$0.747\pm 0.107$	1.00
Bean	70%	$0.959\pm 0.031$	0.30	$0.712\pm 0.120$	0.31

Table 4: Cost-benefit analysis for epistemic gating. Stability is reported as mean

\pm

std of Kendall’s

\tau

for accepted samples. Relative cost

q

follows Eqs. ( 2 )–( 3 ).

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).

Figure 6: PlantVillage: IG and SG saliency maps for clean and noisy low-epistemic (rows 1, 3) and high-epistemic (rows 2, 4) samples. For visualization, heatmaps are lightly Gaussian - smoothed to reduce pixel noise and emphasize structural patterns. Low-epistemic explanations maintain focused attribution on leaf structure and disease areas.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Attribution	Approach
SHAP [ 18 ]	$\phi_{i}=\sum_{S\subseteq\mathcal{F}\setminus\{i\}}\frac{\|S\|!(\|\mathcal{F}\|-\|S\|-1)!}{\|\mathcal{F}\|!}\big[f_{S\cup\{i\}}-f_{S}\big]$	Shapley values
LIME [ 27 ]	$\phi_{i}=w_{i}^{*}$ from $\min_{\mathbf{w}}\sum_{\mathbf{z}}\pi_{\mathbf{x}}(\mathbf{z})(f(\mathbf{z})-\mathbf{w}^{\top}\mathbf{z})^{2}$	Local surrogate
IG [ 33 ]	$\phi_{i}=(x_{i}-x^{\prime}_{i})\int_{0}^{1}\frac{\partial f(\mathbf{x}^{\prime}+\alpha(\mathbf{x}-\mathbf{x}^{\prime}))}{\partial x_{i}}d\alpha$	Path integration
SG/SIG [ 32 ]	$\phi_{i}=\frac{1}{N}\sum_{n=1}^{N}\phi_{i}^{\text{base}}(\mathbf{x}+\epsilon_{n})$	Noise averaging

Appendix

Table 5: Equations of XAI methods used in this work.

Figure 7: Epistemic uncertainty vs. SHAP stability (Kendall’s $\tau$ ) under pooled noise conditions. Green points denote stable explanations ( $\tau\geq 0.7$ ), red points denote unstable. Vertical dashed lines show 30%, 50%, and 70% deferral rates $\nu$ . Marginal distributions illustrate class separation along each axis.

Figure 8: Attribution mass on signal (blue) vs. noise (red) features across epistemic groups (solid: low, hatched: random, cross-hatched: high) and noise feature ratios. Low-epistemic explanations maintain focus on signal features better.

Explainable Artificial Intelligence

Explainable artificial intelligence develops ways to make AI decisions understandable, testable, and useful to people.

See all 326 papers

Concerning Uncertainty -- A Systematic Survey of Uncertainty-Aware XAI
Mar 27, 2026Helena Löfström, Tuwe Löfström, Anders Hjort +1Explainable Artificial Intelligence Uncertainty Quantification
Jönköping International Business School, Gjuterigatan 5, 553 18 Jönköping, Sweden · Jönköping School of Engineering, Gjuterigatan 5, 553 18 Jönköping, Sweden · Eiendomsverdi AS, Nedre Slottsgate 8, 0157 Oslo, Norway +2
Explainable Uncertainty Estimation for Reliable Medical AI
Aug 28, 2026Li Rong Wang, Jamie Duell, Xinran Xu +6Uncertainty Quantification Explainable Artificial Intelligence
College of Computing and Data Science (CCDS) Nanyang Technological University (NTU) A*STAR Centre for Frontier AI Research Singapore · School Of Computing and Digital Technologies Sheffield Hallam University South Yorkshire, England · CCDS NTU Singapore +5
No Single Metric Tells the Whole Story: A Multi-Dimensional Evaluation Framework for Uncertainty Attributions
Mar 25, 2026Emily Schiller, Teodor Chiaburu, Marco Zullich +1Explainability Evaluation Uncertainty Quantification
XITASO GmbH IT & Software Solutions, Augsburg, Germany · The Artificial Intelligence and Cognitive Load Research Lab, University College Cork · Berliner Hochschule für Technik, Berlin, Germany +2

Mar 27, 2026J/K move · Enter open · S save

Concerning Uncertainty -- A Systematic Survey of Uncertainty-Aware XAI

Helena Löfström, Tuwe Löfström, Anders Hjort, Fatima Rabia Yapicioglu

Jönköping International Business School, Gjuterigatan 5, 553 18 Jönköping, Sweden · Jönköping School of Engineering, Gjuterigatan 5, 553 18 Jönköping, Sweden · Eiendomsverdi AS, Nedre Slottsgate 8, 0157 Oslo, Norway+2

Explainable Artificial Intelligence Uncertainty Quantification Explainability Evaluation Uncertainty Propagation

Open paper AskPDF arXiv

This paper surveys uncertainty-aware explainable artificial intelligence (UAXAI), examining how uncertainty is incorporated into explanatory pipelines and how such methods are evaluated. Across the literature, three recurring approaches to uncertainty quantification emerge (Bayesian, Monte Carlo, and Conformal methods), alongside distinct strategies for integrating uncertainty into explanations: assessing trustworthiness, constraining models or explanations, and explicitly communicating uncertainty. Evaluation practices remain fragmented and largely model centered, with limited attention to users and inconsistent reporting of reliability properties (e.g., calibration, coverage, explanation stability). Recent work leans towards calibration, distribution free techniques and recognizes explainer variability as a central concern. We argue that progress in UAXAI requires unified evaluation principles that link uncertainty propagation, robustness, and human decision-making, and highlight counterfactual and calibration approaches as promising avenues for aligning interpretability with reliability.

Abstract

Explainable Artificial Intelligence

Explore similar work

Concerning Uncertainty -- A Systematic Survey of Uncertainty-Aware XAI

Explainable Uncertainty Estimation for Reliable Medical AI

No Single Metric Tells the Whole Story: A Multi-Dimensional Evaluation Framework for Uncertainty Attributions