Research map / Label-efficient and robust learning / Test-time adaptation
Sequential Harmful Shift Detection Without Labels
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 sequential harmful shift detection without production labels by calibrating a learned error estimator's high-error selector and tracking the running proportion of flagged production observations with confidence sequences.
Problem
Detecting distribution shifts that harmfully increase a deployed model's error in sequential production settings when ground-truth labels are unavailable, with valid false alarm control.
Approach
Replace the true error with a secondary error estimator's predictions, calibrate a quantile-based threshold on labeled source data so the selector separates high from low true errors, and track the running proportion of observations exceeding the threshold using a predictably-mixed empirical-Bernstein confidence sequence, rather than the mean estimated error.
Key results
- Quantile detector achieves power 0.83 and FDP 0.11 across synthetic tabular shifts versus 0.67 and 0.41 for the mean detector at epsilon_tol=0.
- On Folktables natural shifts, the quantile detector reaches power 0.48 with FDP 0.019, versus power 0.01 and FDP 0.19 for the mean detector.
- In batch comparison on NHANES I, Detectron reaches power 0.96-0.98 but FDP about 0.60, while SHSD reaches power 0.40-0.74 with FDP 0.02-0.04.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use quantile detection instead of mean detection | SHSD on synthetic tabular shifts · Synthetic tabular shifts, all datasets | power: 0.67 -> 0.83 | Helped |
| use quantile detection instead of mean detection | SHSD on synthetic tabular shifts · Synthetic tabular shifts, all datasets | FDP: 0.41 -> 0.11 | Helped |
| use quantile detection instead of mean detection | SHSD on Folktables · Folktables | power: 0.01 -> 0.48 | Helped |
| use quantile detection instead of mean detection | SHSD on Folktables · Folktables | FDP: 0.19 -> 0.019 | Helped |
| use quantile detection instead of mean detection | SHSD on Folktables · Folktables | mean detection time: 4945 -> 3727 | Helped |
| use SHSD instead of Detectron | Batch detection on NHANESI · NHANESI, OOD size 8593 | FDP: 0.60 -> 0.04 | Helped |
| use SHSD instead of Detectron | Batch detection on NHANESI · NHANESI, OOD size 8593 | power: 0.98 -> 0.74 | Hurt |
| use the first quantile statistic Φq instead of the second statistic Φ²q (ablation) | Quantile detector on Folktables · Folktables | FDP: 0.019 -> 0.004 | Helped |
| use the first quantile statistic Φq instead of the second statistic Φ²q (ablation) | Quantile detector · Synthetic shifts on tabular datasets | power | Hurt |
| use the first quantile statistic Φq instead of the second statistic Φ²q (ablation) | Quantile detector on California · California house prices | shift detection | Hurt |
Surprising findings
- Error estimators are highly imperfect (R-squared mostly below 0.3), yet calibrated thresholds still achieve low FDP in production.
- Detectron raised many false alarms (e.g., 1126 of 1700 at sample size 100) in the batch comparison.
Limitations
- Assumption that false-discovery probability in production does not exceed source level holds only approximately (about 50% of cases violated, typically by small margins).
- Error estimators are often weak (low R-squared), so power depends on estimator accuracy.
- Disagreement-based baseline comparison is limited to batch setting and classification datasets.
Experimental setup
Datasets: California house prices Bike sharing demand HELOC NHANES I CelebA CIFAR-10 Fashion MNIST Folktables
Benchmarks: California house prices Bike sharing demand HELOC NHANES I CelebA Folktables CIFAR-10 Fashion MNIST
Metrics: power false discovery proportion detection time
Baselines: mean detector (Podkopaev and Ramdas plug-in) Detectron
Builds on: Podkopaev and Ramdas 2022 sequential harmful shift detection with labels confidence sequences predictably-mixed empirical-Bernstein
Closest papers by meaning
- Tracking the risk of a deployed model and detecting harmful distribution shifts (ICLR 2022 · similarity 0.92)
- Prediction-Powered Risk Monitoring of Deployed Models for Detecting Harmful Distribution Shifts (ICML 2026 · similarity 0.88)
- Estimating and Explaining Model Performance When Both Covariates and Labels Shift (NeurIPS 2022 · similarity 0.87)
- Learning Label Shift Correction for Test-Agnostic Long-Tailed Recognition (ICML 2024 · similarity 0.84)
- LaSCal: Label-Shift Calibration without target labels (NeurIPS 2024 · similarity 0.83)
- When to retrain a machine learning model (ICML 2025 · similarity 0.83)
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.