Research map / Label-efficient and robust learning / Recognition and long-tail

Suitability Filter: A Statistical Framework for Classifier Evaluation in Real-World Deployment Settings

ICML 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Learning from unlabeled and positive-unlabeled data · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes the suitability filter, a non-inferiority statistical test that uses per-sample suitability signals and a logistic correctness estimator to decide without labels whether a classifier's accuracy on unlabeled user data drops more than a margin below its labeled test accuracy.

Problem

Deployed classifiers may lose accuracy under covariate shift, but ground-truth labels for user data are unavailable, so degradation relative to a labeled test set cannot be directly verified.

Approach

Aggregate model output features (suitability signals such as max softmax, entropy, energy) into per-sample correctness probabilities via a logistic regressor trained on a held-out provider set, then compare the mean estimated correctness on user vs. test data with a one-sided Welch's t-test (non-inferiority test) to output SUITABLE or INCONCLUSIVE, with a theoretical FPR bound under delta-calibrat

Key results

• On FMoW-WILDS ID user data the filter reaches ROC AUC 0.969 and PR AUC 0.967 at m=0, with FPR 0.027 at alpha=0.05.
• On RxRx1-WILDS ID user data the filter attains ROC AUC 1.000 and FPR 0.000; on OOD user data ROC AUC is 0.997.
• On FMoW-WILDS the filter achieves 100% decision accuracy for accuracy drops greater than 3% (about 8.4k of nearly 29k experiments).

Measured effects reported in the paper

Surprising findings

• For accuracy gaps between 0% and 3% the filter yields frequent false negatives, since the evidence is insufficient to reject inferiority at alpha=0.05.

Limitations

• Only detects covariate shift; does not address label shift and assumes only unlabeled target data is available.
• Theoretical false positive guarantee requires delta-calibration that is hard to verify; experiments set the margin correction to zero.
• Limited to classification and accuracy metric; assumes non-adversarial providers and users.

Experimental setup

Datasets: FMoW-WILDS RxRx1-WILDS CivilComments-WILDS

Benchmarks: FMoW-WILDS RxRx1-WILDS CivilComments-WILDS

Metrics: ROC AUC PR AUC accuracy false positive rate at alpha=0.05

Builds on: Welch's t-test dataset inference WILDS benchmark Platt scaling

Closest papers by meaning

• What Does It Take to Build a Performant Selective Classifier? (NeurIPS 2025 · similarity 0.86)
• probly: Uncertainty-Aware Machine Learning (NeurIPS 2026 · similarity 0.84)
• Overcoming Common Flaws in the Evaluation of Selective Classification Systems (NeurIPS 2024 · similarity 0.84)
• SMLE: Safe Machine Learning via Embedded Overapproximation (AAAI 2025 · similarity 0.83)
• Leveraging Uncertainty Estimates To Improve Classifier Performance (ICLR 2024 · similarity 0.82)
• Classification with Conceptual Safeguards (ICLR 2024 · similarity 0.82)