module: difficulty_based¶
-
class
annotlib.difficulty_based.
DifficultyBasedAnnot
(X, y_true, classifiers=None, n_annotators=None, alphas=None, n_splits=5, n_repeats=10, confidence_noise=None, random_state=None)[source]¶ Bases:
annotlib.standard.StandardAnnot
This class implements a simulation technique aiming at quantifying the difficulty of a sample. The estimated difficulty is used in combination with an annotator labelling performance to compute the probability that the corresponding annotator labels the sample correctly.
Parameters: - X: array-like, shape (n_samples, n_features)
Samples of the whole data set.
- y_true: array-like, shape (n_samples)
True class labels of the given samples X.
- n_annotators: int
Number of annotators who are simulated.
- classifiers: sklearn.base.ClassifierMixin | list of ClassifierMixin, shape (n_classifiers)
The classifiers parameter is either a single sklearn classifier supporting :py:method::predict_proba` or a list of such classifiers. If the parameter is not a list, the simplicity scores are estimate by a single classifier, whereas if it is a list, the simplicity scores can be estimated by different classifier types or different parametrisations. The default classifiers parameter is a single SVM
- alphas: array-like, shape (n_annotators)
The entry alphas[a_idx] indicates the annotator labelling performance, which is in the interval (-inf, inf). The following properties are valid: - alphas[a_idx] = 0: annotator with index a_idx makes random guesses, - alphas[a_idx] = inf: annotator with index a_idx is almost always right, - alphas[a_idx] = -inf: annotator with index a_idx is almost always wrong (adversarial).
- n_splits: int
Number of folds of the cross-validation.
- n_repeats: int
Number of repeats of the cross-validation.
- confidence_noise: array-like, shape (n_annotators)
An entry of confidence_noise defines the interval from which the noise is uniformly drawn, e.g. confidence_noise[a] = 0.2 results in sampling n_samples times from U(-0.2, 0.2) and adding this noise to the confidence scores. Zero noise is the default value for each annotator.
- random_state: None | int | instance of :py:class:`numpy.random.RandomState`
The random state used for generating class labels of the annotators.
Examples
>>> from sklearn.datasets import load_iris >>> from sklearn.gaussian_process import GaussianProcessClassifier >>> from sklearn.svm import SVC >>> # load iris data set >>> X, y_true = load_iris(return_X_y=True) >>> # create list of SVM and Gaussian Process classifier >>> classifiers = [SVC(C=1, probability=True, gamma='auto'), SVC(C=3, probability=True), GaussianProcessClassifier()] >>> # set labelling performances of annotators >>> alphas = [-3, 0, 3] >>> # simulate annotators on the iris data set >>> annotators = DifficultyBasedAnnot(X=X, y_true=y_true, classifiers=classifiers, n_annotators=3, alphas=alphas) >>> # the number of annotators must be equal to the number of classifiers >>> annotators.n_annotators() 3 >>> # query class labels of 100 samples from annotators a_0, a_2 >>> annotators.class_labels(X=X[0:100], y_true=y_true[0:100], annotator_ids=[0, 2], query_value=100).shape (100, 3) >>> # check query values >>> annotators.n_queries() array([100, 0, 100]) >>> # query confidence scores of these 100 samples from annotators a_0, a_2 >>> annotators.confidence_scores(X=X[0:100], y_true=y_true[0:100], annotator_ids=[0, 2]).shape (100, 3) >>> # query values are not affected by calling the confidence score method >>> annotators.n_queries() array([100, 0, 100]) >>> # labelling performance of annotator a_0 is adversarial (worse than guessing) >>> annotators.labelling_performance(X=X, y_true=y_true)[0] < 1/len(np.unique(y_true)) True
Attributes: - X_: numpy.ndarray, shape (n_samples, n_features)
Samples of the whole data set.
- Y_: numpy.ndarray, shape (n_samples, n_annotators)
Class labels of the given samples X.
- C_: numpy.ndarray, shape (n_samples, n_annotators)
confidence score for labelling the given samples x.
- C_noise_: numpy.ndarray, shape (n_samples, n_annotators)
The uniformly noise for each annotator and each sample, e.g. C[x_idx, a_idx] indicates the noise for the confidence score of annotator with id a_idx in labelling sample with id x_idx.
- n_annotators_: int
Number of annotators.
- n_queries_: numpy.ndarray, shape (n_annotators)
An entry n_queries_[a] indicates how many queries annotator a has processed.
- queried_flags_: numpy.ndarray, shape (n_samples, n_annotators)
An entry queried_flags_[i, j] is a boolean indicating whether annotator a_i has provided a class label for sample x_j.
- y_true_: numpy.ndarray, shape (n_samples)
The true class labels of the given samples.
- alphas_: array-like, shape (n_annotators)
The entry alphas_[a_idx] indicates the annotator labelling performance, which is in the interval (-inf, inf). The following properties are valid: - alphas_[a_idx] = 0: annotator with index a_idx makes random guesses, - alphas_[a_idx] = inf: annotator with index a_idx is almost always right, - alphas_[a_idx] = -inf: annotator with index a_idx is almost always wrong (adversarial).
- betas_: array-like, shape (n_annotators)
The entry betas_[x_idx] represents the simplicity score of sample X_[x_idx], where betas_[x_idx] is in the interval [0, inf): - betas_[x_idx] = 0: annotator with index a_idx makes random guesses, - betas_[x_idx] = inf: annotator with index a_idx is always right, if alphas_[a_idx] > 0
- n_splits_: int
Number of folds of the cross-validation.
- n_repeats: int
Number of repeats of the cross-validation.
- random_state_: None | int | numpy.random.RandomState
The random state used for generating class labels of the annotators.