Research map / Label-efficient and robust learning / Recognition and long-tail
Missingness Bias Calibration in Feature Attribution Explanations
Shailesh Sridhar, Anton Xue, Eric Wong
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add MCal linear logit calibrator | frozen base classifiers · 8 medical benchmarks (vision, language, tabular) | KL missingness bias: e.g. CheXpert 1.70e-1 -> 8.82e-3; MedQA 1.61e-1 -> 9.44e-4 | Improved |
| use MCal instead of retraining the model on ablated inputs | ViT-B16 · Brain MRI | KL missingness bias: 6.70e-4 (Retrain) vs 7.43e-3 (MCal) | Hurt |
| use MCal instead of retraining | Llama-3.1-8B-Instruct · MedQA | KL missingness bias: 1.70e-1 -> 9.44e-4 | Improved |
| use MCal instead of retraining | XGBoost · PhysioNet, Breast Cancer, CTG | KL missingness bias: Breast Cancer 5.68e-3 -> 1.92e-5 | Mixed |
| use MCal instead of temperature scaling | Llama-3.1-8B-Instruct · MedMCQA | KL missingness bias: 7.81e-1 -> 9.01e-3 | Improved |
| use architecture-native support for missing features (Arch) | XGBoost · Breast Cancer and CTG | KL missingness bias: 1.02e-1 -> 2.13e-1; 1.06e-1 -> 2.85e-1 | Hurt |
| use replacement-based imputation (Replace) | base classifiers · 8 medical benchmarks | KL missingness bias: CheXpert 1.70e-1 -> 9.70e-2; BreakHis 1.87e-1 -> 4.20e-1 | Mixed |
| condition calibrators on ablation rate (ensemble) vs a single unconditioned calibrator (ablation) | MCal · MRI, MedQA, PhysioNet | missingness bias | Improved |
| add MCal (conditioned or unconditioned) | base classifiers · accuracy under ablation | accuracy | Improved |
Closest papers by meaning
- Benchmarking the Attribution Quality of Vision Models (NeurIPS 2024 · similarity 0.85)
- Towards Fine-Grained Interpretability: Counterfactual Explanations for Misclassification with Saliency Partition (CVPR 2025 · similarity 0.85)
- Evaluating Model Bias Requires Characterizing its Mistakes (ICML 2024 · similarity 0.84)
- Prediction models that learn to avoid missing values (ICML 2025 · similarity 0.84)
- Learning What Not to Impute: An Uncertainty-Aware Diffusion Framework for Meaningful Missingness (NeurIPS 2026 · similarity 0.84)
- TRoVe: Discovering Error-Inducing Static Feature Biases in Temporal Vision-Language Models (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
