CSSLClassifier

class scikit_weak.classification.CSSLClassifier
A credal self-supervised learning classifiers as described in “Credal Self-Supervised Learning” by

Julian Lienen and Eyke Huellermeier, NeurIPS 2021. The model iteratively refines its beliefs about the true conditional class probability distribution for weakly labeled instances by maintaining credal sets induced by possibility distributions.

Parameters

  • estimator (estimator class, default=LabelRelaxationNNClassifier) – Base estimator objects to be fitted. Should support predict and predict_proba

  • n_iterations (int, default=10) – The number of iterations for fitting

  • random_state (int, default=None) – Random seed

  • p_data (np.ndarray, default=None) – Optional prior probability distribution of the class frequencies

  • p_hist_buffer_size (int, default=64) – Buffer size of the model prediction history

Variables

  • estimator (estimator) – The last fitted estimator

  • _n_classes (int) – The number of unique classes in y

fit(X, y)

y is list of discrete fuzzy labels

predict(X)

Returns predictions for the given X

predict_proba(X)

Returns probability distributions for the given X

References

[1] Lienen, J., Hüllermeier, E. (2021).

Credal Self-Supervised Learning. Advances in Neural Information Processing Systems, 34.