filter#

Methods to flag which examples have label issues in multi-label classification datasets. Here each example can belong to one or more classes, or none of the classes at all. Unlike in standard multi-class classification, model-predicted class probabilities need not sum to 1 for each row in multi-label classification.

Functions:

`find_label_issues`(labels, pred_probs[, ...])	Identifies potentially mislabeled examples in a multi-label classification dataset.
`find_multilabel_issues_per_class`(labels, ...)	Identifies potentially bad labels for each example and each class in a multi-label classification dataset.

cleanlab.multilabel_classification.filter.find_label_issues(labels, pred_probs, return_indices_ranked_by=None, rank_by_kwargs={}, filter_by='prune_by_noise_rate', frac_noise=1.0, num_to_remove_per_class=None, min_examples_per_class=1, confident_joint=None, n_jobs=None, verbose=False)[source]#

Identifies potentially mislabeled examples in a multi-label classification dataset. An example is flagged as with a label issue if any of the classes appear to be incorrectly annotated for this example.

Parameters:

labels (List[List[int]]) – List of noisy labels for multi-label classification where each example can belong to multiple classes. This is an iterable of iterables where the i-th element of labels corresponds to a list of classes that the i-th example belongs to, according to the original data annotation (e.g. labels = [[1,2],[1],[0],..]). This method will return the indices i where the inner list labels[i] is estimated to have some error. For a dataset with K classes, each class must be represented as an integer in 0, 1, …, K-1 within the labels.
pred_probs (np.ndarray) –
An array of shape (N, K) of model-predicted class probabilities. Each row of this matrix corresponds to an example x and contains the predicted probability that x belongs to each possible class, for each of the K classes (along its columns). The columns need not sum to 1 but must be ordered such that these probabilities correspond to class 0, 1, …, K-1.

Note

Estimated label quality scores are most accurate when they are computed based on out-of-sample pred_probs from your model. To obtain out-of-sample predicted probabilities for every example in your dataset, you can use cross-validation. This is encouraged to get better results.
return_indices_ranked_by ({None, 'self_confidence', 'normalized_margin', 'confidence_weighted_entropy'}, default = None) – This function can return a boolean mask (if None) or an array of the example-indices with issues sorted based on the specified ranking method. Refer to documentation for this argument in filter.find_label_issues for details.
rank_by_kwargs (dict, optional) – Optional keyword arguments to pass into scoring functions for ranking by label quality score (see rank.get_label_quality_scores).
filter_by ({'prune_by_class', 'prune_by_noise_rate', 'both', 'confident_learning', 'predicted_neq_given', 'low_normalized_margin', 'low_self_confidence'}, default 'prune_by_noise_rate') – The specific Confident Learning method to determine precisely which examples have label issues in a dataset. Refer to documentation for this argument in filter.find_label_issues for details.
frac_noise (float, default = 1.0) – This will return the “top” frac_noise * num_label_issues estimated label errors, dependent on the filtering method used, Refer to documentation for this argument in filter.find_label_issues for details.
num_to_remove_per_class (array_like) – An iterable that specifies the number of mislabeled examples to return from each class. Refer to documentation for this argument in filter.find_label_issues for details.
min_examples_per_class (int, default = 1) – The minimum number of examples required per class below which examples from this class will not be flagged as label issues. Refer to documentation for this argument in filter.find_label_issues for details.
confident_joint (np.ndarray, optional) – An array of shape (K, 2, 2) representing a one-vs-rest formatted confident joint, as is appropriate for multi-label classification tasks. Entry (c, i, j) in this array is the number of examples confidently counted into a (class c, noisy label=i, true label=j) bin, where i, j are either 0 or 1 to denote whether this example belongs to class c or not (recall examples can belong to multiple classes in multi-label classification). The confident_joint can be computed using count.compute_confident_joint with multi_label=True. If not provided, it is computed from the given (noisy) labels and pred_probs.
n_jobs (optional) – Number of processing threads used by multiprocessing. Refer to documentation for this argument in filter.find_label_issues for details.
verbose (optional) – If True, prints when multiprocessing happens.

Return type:

ndarray

Returns:

label_issues (np.ndarray) – If return_indices_ranked_by left unspecified, returns a boolean mask for the entire dataset where True represents an example suffering from some label issue and False represents an example that appears accurately labeled.

If return_indices_ranked_by is specified, this method instead returns a list of indices of examples identified with label issues (i.e. those indices where the mask would be True). Indices are sorted by the likelihood that all classes are correctly annotated for the corresponding example.

Note

Obtain the indices of examples with label issues in your dataset by setting return_indices_ranked_by.

cleanlab.multilabel_classification.filter.find_multilabel_issues_per_class(labels, pred_probs, return_indices_ranked_by=None, rank_by_kwargs={}, filter_by='prune_by_noise_rate', frac_noise=1.0, num_to_remove_per_class=None, min_examples_per_class=1, confident_joint=None, n_jobs=None, verbose=False)[source]#

Identifies potentially bad labels for each example and each class in a multi-label classification dataset. Whereas find_label_issues estimates which examples have an erroneous annotation for any class, this method estimates which specific classes are incorrectly annotated as well. This method returns a list of size K, the number of classes in the dataset.

Parameters:

labels (List[List[int]]) – List of noisy labels for multi-label classification where each example can belong to multiple classes. Refer to documentation for this argument in find_label_issues for further details. This method will identify whether labels[i][k] appears correct, for every example i and class k.
pred_probs (np.ndarray) – An array of shape (N, K) of model-predicted class probabilities. Refer to documentation for this argument in find_label_issues for further details.
return_indices_ranked_by ({None, 'self_confidence', 'normalized_margin', 'confidence_weighted_entropy'}, default = None) – This function can return a boolean mask (if this argument is None) or a sorted array of indices based on the specified ranking method (if not None). Refer to documentation for this argument in filter.find_label_issues for details.
rank_by_kwargs (dict, optional) – Optional keyword arguments to pass into scoring functions for ranking by. label quality score (see rank.get_label_quality_scores).
filter_by ({'prune_by_class', 'prune_by_noise_rate', 'both', 'confident_learning', 'predicted_neq_given', 'low_normalized_margin', 'low_self_confidence'}, default = 'prune_by_noise_rate') – The specific method that can be used to filter or prune examples with label issues from a dataset. Refer to documentation for this argument in filter.find_label_issues for details.
frac_noise (float, default = 1.0) – This will return the “top” frac_noise * num_label_issues estimated label errors, dependent on the filtering method used, Refer to documentation for this argument in filter.find_label_issues for details.
num_to_remove_per_class (array_like) – This parameter is an iterable that specifies the number of mislabeled examples to return from each class. Refer to documentation for this argument in filter.find_label_issues for details.
min_examples_per_class (int, default = 1) – The minimum number of examples required per class to avoid flagging as label issues. Refer to documentation for this argument in filter.find_label_issues for details.
confident_joint (np.ndarray, optional) – An array of shape (K, 2, 2) representing a one-vs-rest formatted confident joint. Refer to documentation for this argument in cleanlab.multilabel_classification.filter.find_label_issues for details.
n_jobs (optional) – Number of processing threads used by multiprocessing. Refer to documentation for this argument in filter.find_label_issues for details.
verbose (optional) – If True, prints when multiprocessing happens.

Return type:

Union[ndarray, Tuple[List[ndarray], List[Any], List[ndarray]]]

Returns:

per_class_label_issues (list(np.ndarray)) – By default, this is a list of length K containing the examples where each class appears incorrectly annotated. per_class_label_issues[k] is a Boolean mask of the same length as the dataset, where True values indicate examples where class k appears incorrectly annotated.

For more details, refer to cleanlab.multilabel_classification.filter.find_label_issues.

Otherwise if return_indices_ranked_by is not None, then this method returns 3 objects (each of length K, the number of classes): label_issues_list, labels_list, pred_probs_list.

label_issues_list: an ordered list of indices of examples where class k appears incorrectly annotated, sorted by the likelihood that class k is correctly annotated.
labels_list: a binary one-hot representation of the original labels, useful if you want to compute label quality scores.
pred_probs_list: a one-vs-rest representation of the original predicted probabilities of shape (N, 2), useful if you want to compute label quality scores. pred_probs_list[k][i][0] is the estimated probability that example i belongs to class k, and is equal to: 1 - pred_probs_list[k][i][1].