Missingness Bias Calibration in Feature Attribution Explanations

Shailesh Sridhar; Anton Xue; Eric Wong

cs.LGMar 5, 2026

Missingness Bias Calibration in Feature Attribution Explanations

Authors: Shailesh Sridhar, Anton Xue, Eric Wong

Organizations: University of Pennsylvania · University of Texas at Austin

Abstract

Popular explanation methods often produce unreliable feature importance scores due to missingness bias, a systematic distortion that arises when models are probed with ablated, out-of-distribution inputs. Existing solutions treat this as a deep representational flaw that requires expensive retraining or architectural modifications. In this work, we challenge this assumption and show that missingness bias can be effectively treated as a superficial artifact of the model's output space. We introduce MCal, a lightweight post-hoc method that corrects this bias by fine-tuning a simple linear head on the outputs of a frozen base model. Surprisingly, we find this simple correction consistently reduces missingness bias and is competitive with, or even outperforms, prior heavyweight approaches across diverse medical benchmarks spanning vision, language, and tabular domains.

Figures & tables

Panel of Figure 1 — Figure 1: Removing irrelevant features can cause a misdiagnosis. A fine-tuned ViT ( Dosovitskiy et al., 2020 ) correctly predicts “tumor” on the clean image (left) and a subset of the relevant features (middle). However, masking irrelevant features flips the prediction to “healthy”, despite the tumor remaining visible (right). For visualization, gray stripes denote zero-valued pixels, and images are contrast-boosted.

Figure 2: Feature ablations induce class distribution shifts. Masking non-critical regions skews predictions towards the “healthy” class, even when tumors remain visible. This effect, known as missingness bias , causes the model to misclassify inputs that retain relevant features, and undermines the reliability of feature attribution explanations.

Figure 3: MCal corrects class distribution shifts induced by input ablations. The model initially predicts “healthy” from the ablated input. MCal applies a learned transformation $R_{\theta}$ to adjust the output probabilities, thereby restoring alignment with expected class distributions. This calibration method is model-agnostic, requiring only the classifier’s output probabilities of each class.

Panel of Figure 4 — Figure 4: Geometric intuition of MCal on a synthetic dataset. Missingness bias causes the uncalibrated outputs to shift. For instance, the Class A cluster (blue circles) is pulled towards the Class B vertex, leading to systematic misclassification and low accuracy. MCal applies an optimal affine transformation to the uncalibrated outputs, correcting the shift and improving accuracy.

Figure 5: Calibrated models have better explanations. Compared to an uncalibrated baseline model (Base), LIME and SHAP explanations on MCal-calibrated models have more accurate feature importance scores (sufficiency

\downarrow

). In addition, calibrated models are also more robust to feature ablations (sensitivity

\downarrow

). Results are shown for the MRI dataset using an unconditioned calibrator.

Figure 6: Conditioning on ablation rate improves MCal. Fitting an ensemble of calibrators at a discretized set of ablation rates can help reduce the overall missingness rate, compared to using a single unconditioned calibrator. (Left) MRI, (Middle) MedQA, (Right) PhysioNet.

	Dataset	Base	Replace	Retrain	Arch	TempCal	PlattCal	MCal ( ✓ )
Vision	Brain MRI	$1.18\,\mathrm{e}{-1}$	$1.51\,\mathrm{e}{-1}$	$\mathbf{6.70\,e{-4}}$	$1.40\,\mathrm{e}{-1}$	$1.16\,\mathrm{e}{-1}$	$1.27\,\mathrm{e}{-1}$	$7.43\,\mathrm{e}{-3}$
	CheXpert	$1.70\,\mathrm{e}{-1}$	$9.70\,\mathrm{e}{-2}$	$2.67\,\mathrm{e}{-2}$	$1.50\,\mathrm{e}{-1}$	$1.65\,\mathrm{e}{-1}$	$2.02\,\mathrm{e}{-1}$	$\mathbf{8.82\,e{-3}}$
	BreakHis	$1.87\,\mathrm{e}{-1}$	$4.20\,\mathrm{e}{-1}$	$2.19\,\mathrm{e}{-2}$	$1.54\,\mathrm{e}{-1}$	$1.86\,\mathrm{e}{-1}$	$1.66\,\mathrm{e}{-1}$	$\mathbf{4.29\,e{-3}}$
Language	MedQA	$1.61\,\mathrm{e}{-1}$	$1.50\,\mathrm{e}{-1}$	$1.70\,\mathrm{e}{-1}$	$2.68\,\mathrm{e}{-2}$	$1.57\,\mathrm{e}{-1}$	$9.48\,\mathrm{e}{-2}$	$\mathbf{9.44\,e{-4}}$
Language	MedMCQA	$1.89\,\mathrm{e}{-1}$	$2.59\,\mathrm{e}{-1}$	$1.52\,\mathrm{e}{-1}$	$1.40\,\mathrm{e}{-1}$	$7.81\,\mathrm{e}{-1}$	$1.13\,\mathrm{e}{-1}$	$\mathbf{9.01\,e{-3}}$
Tabular	PhysioNet	$1.17\,\mathrm{e}{-1}$	$1.20\,\mathrm{e}{-1}$	$5.59\,\mathrm{e}{-3}$	$8.14\,\mathrm{e}{-2}$	$1.17\,\mathrm{e}{-1}$	$1.19\,\mathrm{e}{-1}$	$\mathbf{5.01\,e{-3}}$

Table 1: MCal is an effective and lightweight way to reduce missingness bias. It repeatedly outperforms more computationally expensive baselines, such as retraining and architecture modification. We report the KL divergence-based metric in Equation 1 .

Figure 7: MCal does not harm classifier accuracy. Across different ablation levels, the accuracy of uncalibrated vs. calibrated classifiers is comparable. This also holds when the input is clean (ablation fraction zero). (Left) MRI, (Middle) MedQA, (Right) PhysioNet.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 8 — Figure 8: MCal training dynamics and performance on different dataset sizes. For different amounts ( $n$ ) of clean-ablated input pairs, we show the training loss curves and test-set accuracies. Each training run consisted of $5000$ iterations, and all runs finished in $\leq 5$ seconds.

Figure 9: MCal can outperform retrain due to its specialization Fitting an ensemble of calibrators at a discretized set of ablation rates can even outperform retraining the model at times (Left) Breast Cancer, (Middle) CTG, (Right) PhysioNet.

Panel of Figure 10 — Figure 10: For a selected example question, MCal results in different feature importances , for example medically relevant features/terms such as ”hematuria” gain importance in the calibrated heatmap. The model task is, for the above question, to choose between the Options: A: Cyclophosphamide, B: Cisplatin, C: Mesna, D: Bleomycin.

Figure 11: Direct comparison of LIME-derived feature importance values, with (a) Uncalibrated, (b) Calibrated, and (c) demonstrating the difference between the two, i.e., (b) - (a).

Panel of Figure 12 — Figure 12: Selected examples of LIME on the MRI dataset. In calibrated models, we observe that LIME tends to assign less importance to border patches, where relevant features are less likely to occur. The four classes are: Meningioma (0), Glioma (1), Pituitary Tumor (2), and No Tumor (3).

Feature Attribution

Feature attribution estimates which input features or internal components contribute to a model’s predictions.

See all 231 papers

XtrAIn: Training-Guided Occlusion for Feature Attribution
Jun 9, 2026Thodoris Lymperopoulos, Ioannis Kakogeorgiou, Denia KanellopoulouFeature Attribution
NCSR Demokritos, Athens, Greece
What is Missing? Explaining Neurons Activated by Absent Concepts
Mar 10, 2026Robin Hesse, Simone Schaub-Meyer, Janina Hesse +2Feature Visualization Neural Network Interpretability
Max Planck Institute for Informatics, SIC · Department of Computer Science, Technical University of Darmstadt · Leibniz Institute for Resilience Research +2
Attribution Upsampling should Redistribute, Not Interpolate
Mar 17, 2026Vincenzo Buono, Peyman Sheikholharam Mashhadi, Mahmoud Rahat +2Feature Attribution Saliency Map Evaluation
Halmstad University, Halmstad, Sweden

Jun 9, 2026J/K move · Enter open · S save

XtrAIn: Training-Guided Occlusion for Feature Attribution

Thodoris Lymperopoulos, Ioannis Kakogeorgiou, Denia Kanellopoulou

NCSR Demokritos, Athens, Greece

Feature Attribution

Open paper AskPDF arXiv

Occlusion-based attribution methods provide an intuitive way to estimate feature importance by perturbing input features and measuring the resulting change in model output. However, their reliability is strongly affected by how feature removal is implemented: externally selected baselines can introduce bias, out-of-distribution samples, and unstable explanations, while in nonlinear models the occlusion of a set of features can also alter the contribution of non-occluded features. We refer to this effect as attribution shift, as the attribution scores of the non-occluded features drift from their initial values. To challenge these major issues that render explanations unstable, we introduce XtrAIn, a training-guided attribution method that transfers the occlusion operation from the input space to the parameter space. Instead of replacing input values with hand-crafted baselines, XtrAIn follows the model's training trajectory and measures how feature-associated parameter updates affect the output logits. We further introduce Xstep, a lightweight approximation for reducing computational cost, and XtrAIn+, a target-focused variant that emphasizes updates aligned with the target class. Experiments on controlled image datasets and PAM50 breast-cancer subtype classification show that the proposed methods produce cleaner and more interpretable attribution patterns than standard attribution baselines. Overall, XtrAIn provides a training-aware perspective on feature attribution and offers a useful diagnostic tool for studying how feature-level evidence is formed during training.

Abstract

Feature Attribution

Explore similar work

XtrAIn: Training-Guided Occlusion for Feature Attribution

What is Missing? Explaining Neurons Activated by Absent Concepts

Attribution Upsampling should Redistribute, Not Interpolate