Noise models and filtering families

noislearn focuses on classification settings where the training data may contain wrong labels, ambiguous instances, or local inconsistencies.

Main noise families

Family	Main idea	Typical use
Noise generators	Inject controlled label corruption with URLF or class-dependent NAR	Reproducible benchmark setup
Distance-based	Compare each instance with its local neighborhood	Fast baselines and local consistency checks
Classifier-based	Compare observed labels against out-of-fold predictions	Label-noise detection from model disagreement
Filter ensemble	Aggregate several filters by votes or scores	Robust decisions across heterogeneous criteria
Noise-score filter	Threshold a continuous suspiciousness score	Controlled removal from ranking outputs
TabPFN-based	Use TabPFN as the base learner and inspect its local explanations	Strong tabular classifier plus explanation support
CNC-NOS	Combine several filters and a weighted noise score	Higher-level cleaning pipeline

Label noise

Label noise appears when an instance is assigned the wrong class. This is the main target of the filtering algorithms in the repository.

Attribute noise

Attribute noise affects feature values rather than the target label. The current API is primarily centered on class-noise filtering, not on feature corruption repair.

Why filtering can be useful

• It can remove or relabel suspicious samples before training the final classifier.
• It can improve downstream generalization when the training labels are unreliable.
• It can provide a structured report that helps audit what was removed and why.

When each family is useful

• Use distance-based filters when you want a simple local consistency baseline.
• Use classifier-based filters when you want out-of-fold disagreement logic.
• Use FilterEnsembleFilter when you want to aggregate several detectors and inspect support, vote, and score matrices.
• Use NoiseScoreFilter when you already have a noise_score_ vector and want a mean, quantile, adaptive, or rational-valley threshold.
• Use TabPFN-based filters when you want stronger tabular predictions and local SHAP reports.
• Use CNC-NOS when you want a higher-level cleaning strategy that combines several filters.

Detection reports

Filters that support the shared detection interface expose noisy_mask_, noisy_indices_, noisy_fraction_, detection_report_, and get_detection_report(). When available, noise_score_ stores a continuous suspiciousness score where larger values indicate more suspicious samples.

Use action="remove" to return only kept samples from fit_resample. Use action="detect" to keep the input unchanged and inspect the report.