Skip to content

Software Quality

Label-error detection

Finding the examples in a labeled dataset whose label a model trained on the rest of the data disputes.

Label-error detection finds the examples in a labeled dataset whose label is probably wrong, so that a person can review them before the dataset trains or evaluates a model. A model is trained on the data, and each example's label is compared with the model's prediction for that example made while it was held out of training. An example the model confidently places in a different class is a candidate error. The labels are the expected values a learned model is held to: a wrong label in training data teaches the error, and a wrong label in a test set scores a correct prediction as a mistake.

The predictions have to be out-of-sample. A model scoring the examples it was trained on has fitted their labels, the wrong ones included, so it tends to agree with them. Tools obtain out-of-sample predictions by cross-validation: train on all folds but one, and predict the held-out fold.

How a disagreement becomes a flag

  • Confident learning compares each example's predicted probability for each class with a per-class threshold, counts the examples that confidently belong to a class other than their label, and from those counts estimates the joint distribution of given and true labels (Northcutt, Jiang, et al. 2021)1.
  • Ensemble filters train several different classifiers and flag an example that all of them, or a majority, misclassify.
  • Training dynamics use how a model's margin on each example evolves during training: the area under the margin statistic separates mislabeled examples from clean ones by the difference in those dynamics (Pleiss et al. 2020)2.

Confident learning assumes that whether a label is wrong depends only on the example's true class, not on the example itself, so that a leopard is more likely to be mislabeled a jaguar than a bathtub (Northcutt, Jiang, et al. 2021)1. Every method shares a deeper limit: the model learns from the same labels it audits, so an error annotators made consistently can become what the model predicts, and is then not flagged.

What it catches

  • Wrong labels. An example annotated with a class it does not belong to, through a slip, a misread instruction, or an annotator guessing.
  • Ambiguous examples and overlapping classes. An example that fits two classes, or two class names that cover the same things, shows up as a cluster of disagreements between the same pair of classes.
  • Mislabeled mined data. Categories taken from tools rather than assigned by annotators, such as the type field of an issue report, carry the errors of whoever filled them in.
  • Missing and misplaced annotations in object detection, where a label is a box: an object nobody boxed, or a box with the wrong class or position.

What it does not catch:

  • an error the model reproduces because annotators made it consistently;
  • whether a flagged example is really wrong, which takes a person or a second annotation;
  • a flawed labeling scheme, where the classes themselves are the wrong ones for the task.

Tools

  • cleanlab — Python library that finds label issues from a model's predicted probabilities, for classification, multi-label, token classification, regression, image segmentation, object detection, and data labeled by several annotators.
  • doubtlab — combines "reasons" for doubt, such as low confidence, a wrong prediction or disagreement between models, into a ranked list of examples to recheck.
  • nessie — research implementation of 18 annotation-error detection methods for text classification and sequence labeling.

When to use, when not

Use:

  • On a test set before models are compared on it, since its errors decide part of the ranking.
  • On crowdsourced labels and on categories mined from tools, where nobody checked each label.
  • To order a relabeling budget: reviewers start with the examples the detector ranks most suspect.

Don't:

  • To delete or relabel flagged examples without review. A correct but unusual example is flagged as well, and removing it narrows the data the model learns.
  • With in-sample predictions, which agree with the labels the model memorized.
  • On data so noisy that the model cannot learn the classes. A filter built from that model cannot be accurate (Brodley and Friedl 1999)3.

Evidence

  • Benchmark test sets carry errors that change model rankings. Across the test sets of 10 commonly used vision, text and audio datasets, at least 3.3% of labels were wrong on average, and at least 6% of the ImageNet validation set. Crowdsourced review confirmed 51% of the candidates confident learning flagged, on average across the datasets. On ImageNet with corrected labels, ResNet-18 outperforms ResNet-50 once the share of originally mislabeled test examples rises by 6% (Northcutt, Athalye, et al. 2021)4.
  • Mined software-engineering data is mislabeled too. In a manual examination of more than 7,000 issue reports from five open-source projects, 33.8% of reports filed as bugs did not refer to a code fix, and on average 39% of files marked as defective had never had a bug (Herzig et al. 2013)5.
  • Ensembles and aggregated scores detect best across tasks. Reimplementing 18 detection methods on 9 English datasets, Klie and colleagues found ensemble and aggregation methods (Borda count, diverse ensemble, label aggregation, retag) worked best overall, and that skipping cross-validation substantially reduced the recall of model-based methods (Klie et al. 2023)6.
  • A simple loss ranking beats elaborate detectors on real errors. Ranking out-of-sample examples by the loss of a fine-tuned pre-trained language model outperformed more complex detectors by 9–36 points of area under the precision-recall curve on real errors in three datasets. Realistic, human-originated label noise was harder to detect than synthetic noise (Chong et al. 2022)7.
  • Filters trade discarded good data against retained bad data. In ensemble filtering on five datasets, consensus filters rarely discarded good data but kept mislabeled data: at 40% noise they retained a mislabeled example with a probability of 50% to 76%. Majority filters balanced the two errors more evenly (Brodley and Friedl 1999)3.
  • Removing flagged training data lowers test error. On WebVision50, removing the 17.8% of training data flagged by the area under the margin reduced test error from 21.4% to 19.8% (Pleiss et al. 2020)2.
  • Detection carries over to object detection. Among the 200 highest-loss proposals per dataset, a loss-based detector found real label errors with a precision of up to 71.5% on commonly used public test sets and 97% on a proprietary production dataset (Schubert et al. 2024)8.

Classification

  • Data-quality characteristics: Accuracy (a wrong label teaches a model the wrong answer, and in a test set marks a correct prediction as an error).
  • Area: Labeled datasets that train or evaluate a model — annotated images, text and audio, crowdsourced labels, and categories mined from tools such as issue trackers.
  • Guarantee: Empirical — a flag is a candidate for a person to review, not a verdict; an unflagged label may still be wrong, especially where the model learned the same mistake the annotators made.

Referenced by

References

  1. Northcutt, Curtis G., Lu Jiang, and Isaac L. Chuang. 2021. "Confident Learning: Estimating Uncertainty in Dataset Labels." Journal of Artificial Intelligence Research 70. https://doi.org/10.1613/jair.1.12125. ↩↩

  2. Pleiss, Geoff, Tianyi Zhang, Ethan R. Elenberg, and Kilian Q. Weinberger. 2020. "Identifying Mislabeled Data using the Area Under the Margin Ranking." Advances in Neural Information Processing Systems 33 (NeurIPS 2020). https://arxiv.org/pdf/2001.10528. ↩↩

  3. Brodley, Carla E., and Mark A. Friedl. 1999. "Identifying Mislabeled Training Data." Journal of Artificial Intelligence Research 11. https://doi.org/10.1613/jair.606. ↩↩

  4. Northcutt, Curtis G., Anish Athalye, and Jonas Mueller. 2021. "Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks." Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks. https://arxiv.org/pdf/2103.14749. ↩

  5. Herzig, Kim, Sascha Just, and Andreas Zeller. 2013. "It's not a bug, it's a feature: How misclassification impacts bug prediction." Proceedings of the 35th International Conference on Software Engineering (ICSE 2013). https://doi.org/10.1109/icse.2013.6606585. ↩

  6. Klie, Jan-Christoph, Bonnie Webber, and Iryna Gurevych. 2023. "Annotation Error Detection: Analyzing the Past and Present for a More Coherent Future." Computational Linguistics 49 (1). https://doi.org/10.1162/coli_a_00464. ↩

  7. Chong, Derek, Jenny Hong, and Christopher D. Manning. 2022. "Detecting Label Errors by using Pre-Trained Language Models." Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP 2022). https://doi.org/10.18653/v1/2022.emnlp-main.618. ↩

  8. Schubert, Marius, Tobias Riedlinger, Karsten Kahl, et al. 2024. "Identifying Label Errors in Object Detection Datasets by Loss Inspection." Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2024). https://openaccess.thecvf.com/content/WACV2024/papers/Schubert_Identifying_Label_Errors_in_Object_Detection_Datasets_by_Loss_Inspection_WACV_2024_paper.pdf. ↩