labelnoise

labelnoise finds examples in classification data whose given label is probably wrong. It works with any classifier: all it needs are the labels and out-of-sample predicted probabilities, for example from cross-validation.

It implements confident learning (Northcutt, Jiang and Chuang, 2021):

per-class confidence thresholds and the confident joint of given and true labels;
calibrated estimates of the joint distribution and the noise transition matrices;
the five rank-and-prune rules for flagging label issues;
model-agnostic label quality scores (Kuan and Mueller, 2022);
a class-level health summary that shows which classes are most often confused.

Installation

Install the released version from CRAN:

install.packages("labelnoise")

Or the development version from GitHub:

# install.packages("remotes")
remotes::install_github("ridsonap/labelnoise")

Example

library(labelnoise)

# Simulated data with known errors, in the yardstick convention
sim <- simulate_label_noise(n = 1000, noise_rate = 0.1, signal = 3, seed = 1)

issues <- find_label_issues(sim, truth = given)
flagged <- issues[issues$is_label_issue, ]
head(flagged[order(flagged$label_quality), ])

health_summary(sim, truth = given)

See the documentation site for the full reference and the introductory vignette.

Works with your modelling framework

Every function accepts the same inputs:

Input	Example
Labels and a probability matrix	`find_label_issues(labels, pred_probs = probs)`
Data frame of predictions (`.pred_*` columns)	`find_label_issues(preds, truth = class)`
tidymodels resampling results	`find_label_issues(fit_resamples(..., control = control_resamples(save_pred = TRUE)))`
mlr3	`find_label_issues(resample(task, lrn("classif.rpart", predict_type = "prob"), rsmp("cv")))`
caret	`find_label_issues(train(..., trControl = trainControl(classProbs = TRUE, savePredictions = "final")))`

Results come back as a tibble with one row per example, in the original row order, so they can be joined back to the data for review.

Caveats

Probabilities must be out-of-sample; in-sample predictions hide errors.
Results are only as good as the classifier. With a weak model, correct labels get flagged and the number of errors is overestimated.
A flag is a reason to look, not a verdict.

Relation to cleanlab

labelnoise is an independent implementation in R from the published papers, not a port of the Python package cleanlab. Its tests compare results with cleanlab 2.9.0, run as a black box on simulated data: the confident joint, all five filtering rules, the estimated number of issues and the self-confidence and normalised-margin scores agree exactly.

References

Northcutt, C., Jiang, L., & Chuang, I. (2021). Confident learning: Estimating uncertainty in dataset labels. Journal of Artificial Intelligence Research, 70, 1373–1411. https://doi.org/10.1613/jair.1.12125

Kuan, J., & Mueller, J. (2022). Model-agnostic label quality scoring to detect real-world label errors. ICML DataPerf Workshop. https://doi.org/10.5281/zenodo.18168246