Co-authored-by: Olivier Grisel <olivier.grisel@ensta.org>
Co-authored-by: Julien Jerphanion <git@jjerphan.xyz>
Co-authored-by: Thomas J. Fan <thomasjpfan@gmail.com>
* resolve ambiguity of the nested cross-val example
* Update examples/model_selection/plot_nested_cross_validation_iris.py
make it more explict
Co-authored-by: Olivier Grisel <olivier.grisel@ensta.org>
Co-authored-by: Olivier Grisel <olivier.grisel@ensta.org>
* Initial implementation
* Forgot to add to second __add__ list
* Update split method parameter doc
* Added example; changed default test_size to 0.1; added to author list
* StratifiedGroupKFold impl and other improvements
* Add class to __all__ spec
* Remove random_state when no shuffle
* Tighter formatting
* Update the implementation of StratifiedGroupKFold
* Add StratifiedGroupKFold to __init__
* Add y checks to StartifiedGroupKFold
* Raise error if n_splits > max num samples in class
* Warn if n_splits > mn num samples in class
* Add SGKfold to general repr test
* Add SGKFold to 2d_y test case
* Add SGKfold to value erros test case
Parameters are the same as for StratifiedKFold
to ensure similar behavior given n_groups == n_samples
* Add SGKFold to StratifiedKFold test cases
The idea is to ensure similar behavior when groups are trivial
(n_groups == n_samples)
* Add SGKFold to reproducibility test case
* Add SGKFold to GroupKFold test case
* Add SGKFold to nested cv test case
* Add SGKFold to random_state with shuffle=False test case
* Add SGKFold to constant splits test case
* Fix repr test case
* Fix formatting issues
* Add samples to a fold with least num samples
Required to produce balanced size folds when the distribution of y is
more or less the same
* Remove GroupShuffleSplit impl
* Add notes to StratifiedGroupKFold
* Fix doctest
* Added stratified group kfold tests
* Better variable naming
* Add section to documentation
* Remove leftover StratifiedGroupShuffleSplit import
* Add changelist and reference to original kernel
* Better naming for least populated class check
* Better expression for number of labels
* Remove use of Counter
We already have this data in output of np.unique
* Add tests for homogeneous groups
* Add StratifiedGroupKFold test against GroupKFold
* Add changes to changelist in docstring
* Add StratifiedGroupKFold to classes.rst
* Fix description of StratifiedGroupKFold
* Move license notice out of docstring
* Disambiguate labels to classes in doc
* Add changelog entry
* Fix changelog author entry
* Fix StratifiedGroupKFold docstring
* Better variable names
* Remove defaultdict in favor of numpy indexing
* Extracted best_fold search into a separate method
* Make use of numpy broadcasting instead of for loop
* Encode groups and use arrays instead of dicts
* Use numpy sort instead of python
* Clarify shuffling behavior of StratifiedGroupKF in docs
* Switch name from label_idx to class_idx
* Remove accidentally leftover comment
* Fix np.sort keyword to support numpy < 1.15
* Fix typo in docstring
* Add StratifiedGroupKFold to visualization doc
* Add visualization for uneven group as an example
* Fix image numbers to match updated example
* Add author
* Add SGKF visualization to docs
* Add comments for groups in stratified CV tests
Co-authored-by: Leandro Hermida <hermidal@cs.umd.edu>
Co-authored-by: marrodion <rodion_martynov@epam.com>
Guillaume Lemaitre