Classifier-based filters

These filters compare the observed labels against predictions obtained from one or more base classifiers.

Overview

• ClassificationFilter uses a single classifier and out-of-fold predictions.
• CVCFFilter aggregates fold-wise committee votes.
• FilterEnsembleFilter (FEF) combines several existing filters by vote and score.
• EnsembleFiltering compares several estimators.
• INFFC_old_wrong iteratively fuses a heterogeneous committee.
• IterativePartitioningFilter repeatedly partitions the data and checks agreement.

ClassificationFilter

Bases: BaseEstimator

Cross-validated single-classifier noise filter.

Parameters:

Name	Type	Description	Default
estimator	estimator	Base learner cloned and trained on each fold.	required
cv	int	Number of stratified folds used to generate out-of-fold predictions.	10
action	(remove, detect)	Whether noisy samples are dropped or only detected.	"remove"
random_state	int	Seed used by the stratified splitter.	33

Notes

A sample is flagged as noisy when its out-of-fold prediction differs from the observed label.

fit(X, y)

Fit the filter and cache out-of-fold predictions.

fit_resample(X, y)

Fit the filter and return the filtered or detected data.

get_filter_report()

Return a dictionary with the main fit diagnostics.

get_detection_report()

Return the stored detection report.

Summary of a single-classifier noise filtering run.

CVCFFilter

Bases: BaseEstimator

Cross-validated committees noise filter.

Parameters:

Name	Type	Description	Default
estimator	estimator	Base learner cloned for each fold of the committee.	c45_like
cv	int	Number of stratified folds used to build the committee.	10
vote_rule	(threshold, majority, consensus)	Rule used to flag samples as noisy from the fold disagreements.	"threshold"
threshold	float	Minimum fraction of disagreeing folds required when vote_rule="threshold".	0.5
action	(remove, detect)	Whether noisy samples are dropped or only detected.	"remove"
random_state	int	Seed used by the stratified splitter.	33

Notes

Relabel is not implemented yet.

fit(X, y)

Fit the filter and cache fold-wise predictions and agreement scores.

fit_resample(X, y)

Fit the filter and return the filtered data.

get_filter_report()

Return a dictionary with the main fit diagnostics.

get_detection_report()

Return the stored detection report.

Summary of a cross-validated committees filtering run.

EnsembleFiltering

Bases: BaseEstimator

Ensemble-based noise filter using multiple classifiers.

Parameters:

Name	Type	Description	Default
estimators	sequence of estimators	Base learners combined in the ensemble committee.	required
cv	int	Number of stratified folds used to compute out-of-fold predictions.	10
mode	str	Decision rule used to flag samples as noisy. The current implementation accepts "threshold" and "consensus"; the signature default is kept for compatibility.	"S"
threshold	float	Minimum fraction of disagreeing estimators required when mode="threshold".	0.5
action	(remove, detect)	Whether noisy samples are dropped or only detected.	"remove"
random_state	int	Seed used by the stratified splitter.	33
return_noisy_samples	bool	Stored on the instance for compatibility; the current implementation does not branch on it.	False

Notes

A sample is flagged as noisy when enough estimators disagree with its observed label.

fit(X, y)

Fit the filter and cache ensemble disagreement counts.

fit_resample(X, y)

Fit the filter and return the filtered data.

get_filter_report()

Return a dictionary with the main fit diagnostics.

get_detection_report()

Return the stored detection report.

Summary of an ensemble-based noise filtering run.

FilterEnsembleFilter

FEF accepts either filter instances or (name, filter) pairs. It supports union, majority, consensus, k_of_m, threshold, and weighted_threshold strategies. The report includes per-filter votes, score matrices, normalized weights, support counts, support fractions, base reports, and class-protection metadata.

Bases: BaseEstimator

Combine several noise filters using vote and score aggregation.

Parameters:

Name	Type	Description	Default
base_filters	list	Sequence of fitted or unfitted filters. Each item can be either a filter instance or a (name, filter) pair.	required
strategy	str	Decision rule used to flag samples as noisy. Supported values are "union", "majority", "consensus", "k_of_m", "threshold", and "weighted_threshold".	"majority"
min_votes	int or None	Minimum number of votes required when strategy="k_of_m".	None
vote_threshold	float	Fraction of filters that must vote noisy when strategy="threshold".	0.5
score_threshold	float	Minimum aggregated ensemble score when strategy="weighted_threshold".	0.5
filter_weights	list of float or None	Optional weight per base filter. If omitted, all filters receive equal weight. The weights are normalized internally.	None
refit_base_filters	bool	Whether to clone and fit the base filters inside :meth:fit. When False, the provided filters are treated as already fitted.	True
use_filter_scores	bool	If True, use each base filter's noise_score when available. Otherwise every filter contributes only its binary vote.	True
score_fallback	str	Fallback strategy used when a base filter does not provide noise_score. Only "binary" is implemented.	"binary"
normalize_scores	bool	Whether to normalize each base filter score to [0, 1] before aggregation.	False
min_class_count	int	Minimum number of samples to keep per class after filtering.	2
action	str	Post-fit action. "remove" and "detect" are executed now. "weight" and "relabel" are reserved for future support.	"remove"
random_state	int or None	Optional seed propagated to cloned base filters when they expose a random_state parameter set to None.	None
n_jobs	int or None	Reserved for future parallel execution.	None

Notes

The final detection report preserves the common structure used by the rest of the library and adds ensemble-specific diagnostics such as the vote and score matrices.

fit(X, y)

Fit the ensemble filter and cache the detection report.

fit_resample(X, y)

Fit the filter and return the filtered data.

fit_filter(X, y)

Fit the filter and return the filtered data plus the report.

get_support_matrix()

Return the binary vote matrix of the base filters.

get_score_matrix()

Return the per-filter score matrix used by the ensemble.

get_ensemble_score()

Return the aggregated ensemble noise score.

get_sample_weight(mode='linear')

Return instance weights derived from the ensemble noise score.

get_filter_report()

Return a compact summary of the ensemble run.

get_detection_report()

Return the stored detection report.

Summary of a filter-ensemble noise filtering run.

NoiseScoreFilter

NoiseScoreFilter thresholds an existing score vector or the noise_score_ exposed by another fitted filter. It supports fixed numeric thresholds, mean, quantile, beta-adaptive thresholding, and rational-valley thresholding with a detailed threshold_report_.

Bases: BaseSampler

Filter samples using a threshold on their noise scores.

Parameters:

Name	Type	Description	Default
sampling_strategy	str or dict	Passed to :class:imblearn.base.BaseSampler.	"auto"
noise_filter	object or None	Existing filter exposing a noise_score_ attribute. When fit_filter=True and the filter is not already fitted, it is trained during :meth:fit to obtain those scores.	None
noise_scores	array - like or None	Precomputed noise scores. Values are clipped to [0, 1] during fit.	None
fit_filter	bool	Whether to fit noise_filter inside :meth:fit when it does not yet expose noise_score_.	False
threshold	(mean, quantile, adaptative, rational_valley, adaptative_quadratic)	Thresholding rule. "mean" uses the average noise score. "quantile" uses the requested quantile of the score distribution. "adaptative" fits two beta-shaped score populations and uses their intersection as the threshold. "rational_valley" fits a rational surrogate to a density estimate and only accepts internal minima that are stable enough. Numeric values are used directly.	"mean"
quantile	float or None	Quantile used when threshold="quantile".	None

Notes

The adaptive branch stores overlap_ as the area shared by the two fitted densities. Lower values mean a cleaner separation between the low-score and high-score regions.

fit(X, y)

Cache the noise scores, optionally fit the nested filter and compute the threshold.

fit_resample(X, y)

Fit the filter and return the cleaned sample set.

INFFC_old_wrong

Bases: BaseEstimator

Legacy iterative fusion-of-classifiers noise filter.

Parameters:

Name	Type	Description	Default
estimators	sequence of estimators or None	Base learners used by the committee. If None, a default trio of C4.5-like tree, 1-NN and LDA is used.	None
cv	int	Number of stratified folds used inside each iteration.	10
decision_rule	(majority, consensus, threshold)	Rule used to flag a sample as noisy from the committee disagreements.	"majority"
threshold	float	Minimum disagreement fraction required when decision_rule="threshold".	0.5
action	(remove, detect)	Whether noisy samples are dropped or only detected.	"remove"
max_iter	int	Maximum number of cleaning iterations.	20
max_removed_frac	float	Stop once this fraction of the original training set has been removed.	0.5
random_state	int	Seed used by the stratified splitter in each iteration.	33

Notes

Relabel is not implemented yet.

fit(X, y)

Fit the filter and iteratively remove noisy samples.

fit_resample(X, y)

Fit the filter and return the filtered or detected data.

get_filter_report()

Return a dictionary with the main fit diagnostics.

get_detection_report()

Return the stored detection report.

Summary of a legacy INFFC filtering run.

Per-iteration diagnostics for the legacy INFFC variant.

IterativePartitioningFilter

Bases: BaseEstimator

Iterative partitioning noise filter.

Parameters:

Name	Type	Description	Default
estimator	estimator	Base learner fitted on each partition.	c45_like
n_partitions	int	Number of stratified partitions built at each iteration.	10
vote_rule	(majority, consensus)	Rule used to flag a sample as noisy from the partition disagreements.	"majority"
action	(remove, detect)	Whether noisy samples are dropped or only detected.	"remove"
p_stop	float	Patience threshold expressed as a fraction of the original dataset.	0.01
k_patience	int	Number of consecutive low-yield iterations tolerated before stopping.	3
max_iter	int	Maximum number of cleaning iterations.	20
random_state	int	Seed used by the stratified splitter in each iteration.	33

Notes

Relabel is not implemented yet.

fit(X, y)

Fit the filter and iteratively partition the training data.

fit_resample(X, y)

Fit the filter and return the filtered or detected data.

get_filter_report()

Return a dictionary with the main fit diagnostics.

get_detection_report()

Return the stored detection report.

Summary of an iterative partitioning filtering run.

Per-iteration diagnostics for iterative partitioning filtering.