Version 1.7

Legend for changelogs

• Major Feature something big that you couldn’t do before.

• Feature something that you couldn’t do before.

• Efficiency an existing feature now may not require as much computation or memory.

• Enhancement a miscellaneous minor improvement.

• Fix something that previously didn’t work as documented – or according to reasonable expectations – should now work.

• API Change you will need to change your code to have the same effect in the future; or a feature will be removed in the future.

Version 1.7.dev0

November 2024

Changes impacting many modules

• Enhancement __sklearn_tags__ was introduced for setting tags in estimators. More details in Estimator Tags. By Thomas Fan and Adrin Jalali #29677

• Enhancement Scikit-learn classes and functions can be used while only having a import sklearn import line. For example, import sklearn; sklearn.svm.SVC() now works. By Thomas Fan #29793

• Fix Classes metrics.ConfusionMatrixDisplay, metrics.RocCurveDisplay, calibration.CalibrationDisplay, metrics.PrecisionRecallDisplay, metrics.PredictionErrorDisplay and inspection.PartialDependenceDisplay now properly handle Matplotlib aliases for style parameters (e.g., c and color, ls and linestyle, etc). By Joseph Barbier #30023

• API Change utils.validation.validate_data is introduced and replaces previously private base.BaseEstimator._validate_data method. This is intended for third party estimator developers, who should use this function in most cases instead of utils.check_array and utils.check_X_y. By Adrin Jalali #29696

Support for Array API

Additional estimators and functions have been updated to include support for all Array API compliant inputs.

See Array API support (experimental) for more details.

• Feature model_selection.GridSearchCV, model_selection.RandomizedSearchCV, model_selection.HalvingGridSearchCV and model_selection.HalvingRandomSearchCV now support Array API compatible inputs when their base estimators do. By Tim Head and Olivier Grisel #27096

• Feature preprocessing.LabelEncoder now supports Array API compatible inputs. By Omar Salman #27381

• Feature sklearn.metrics.mean_absolute_error now supports Array API compatible inputs. By Edoardo Abati #27736

• Feature sklearn.metrics.mean_tweedie_deviance now supports Array API compatible inputs. By Thomas Li #28106

• Feature sklearn.metrics.pairwise.cosine_similarity now supports Array API compatible inputs. By Edoardo Abati #29014

• Feature sklearn.metrics.pairwise.paired_cosine_distances now supports Array API compatible inputs. By Edoardo Abati #29112

• Feature sklearn.metrics.cluster.entropy now supports Array API compatible inputs. By Yaroslav Korobko #29141

• Feature sklearn.metrics.mean_squared_error now supports Array API compatible inputs. By Yaroslav Korobko #29142

• Feature sklearn.metrics.pairwise.additive_chi2_kernel now supports Array API compatible inputs. By Yaroslav Korobko #29144

• Feature sklearn.metrics.d2_tweedie_score now supports Array API compatible inputs. By Emily Chen #29207

• Feature sklearn.metrics.max_error now supports Array API compatible inputs. By Edoardo Abati #29212

• Feature sklearn.metrics.mean_poisson_deviance now supports Array API compatible inputs. By Emily Chen #29227

• Feature sklearn.metrics.mean_gamma_deviance now supports Array API compatible inputs. By Emily Chen #29239

• Feature sklearn.metrics.pairwise.cosine_distances now supports Array API compatible inputs. By Emily Chen #29265

• Feature sklearn.metrics.pairwise.chi2_kernel now supports Array API compatible inputs. By Yaroslav Korobko #29267

• Feature sklearn.metrics.mean_absolute_percentage_error now supports Array API compatible inputs. By Emily Chen #29300

• Feature sklearn.metrics.pairwise.paired_euclidean_distances now supports Array API compatible inputs. By Emily Chen #29389

• Feature sklearn.metrics.pairwise.euclidean_distances and sklearn.metrics.pairwise.rbf_kernel now supports Array API compatible inputs. By Omar Salman #29433

• Feature sklearn.metrics.pairwise.linear_kernel, sklearn.metrics.pairwise.sigmoid_kernel, and sklearn.metrics.pairwise.polynomial_kernel now supports Array API compatible inputs. By Omar Salman #29475

• Feature sklearn.metrics.mean_squared_log_error and sklearn.metrics.root_mean_squared_log_error now supports Array API compatible inputs. By Virgil Chan #29709

• Feature preprocessing.MinMaxScaler with clip=True now supports Array API compatible inputs. By Shreekant Nandiyawar #29751

• Support for the soon to be deprecated cupy.array_api module has been removed in favor of directly supporting the top level cupy module, possibly via the array_api_compat.cupy compatibility wrapper. By Olivier Grisel #29639

Metadata routing

Refer to the Metadata Routing User Guide for more details.

• Feature semi_supervised.SelfTrainingClassifier now supports metadata routing. The fit method now accepts **fit_params which are passed to the underlying estimators via their fit methods. In addition, the predict, predict_proba, predict_log_proba, score and decision_function methods also accept **params which are passed to the underlying estimators via their respective methods. By Adam Li #28494

• Feature ensemble.StackingClassifier and ensemble.StackingRegressor now support metadata routing and pass **fit_params to the underlying estimators via their fit methods. By Stefanie Senger #28701

• Feature model_selection.learning_curve now supports metadata routing for the fit method of its estimator and for its underlying CV splitter and scorer. By Stefanie Senger #28975

• Feature compose.TransformedTargetRegressor now supports metadata routing in its fit and predict methods and routes the corresponding params to the underlying regressor. By Omar Salman #29136

• Feature feature_selection.SequentialFeatureSelector now supports metadata routing in its fit method and passes the corresponding params to the model_selection.cross_val_score function. By Omar Salman #29260

• Feature model_selection.permutation_test_score now supports metadata routing for the fit method of its estimator and for its underlying CV splitter and scorer. By Adam Li #29266

• Feature feature_selection.RFE and feature_selection.RFECV now support metadata routing. By Omar Salman #29312

• Feature model_selection.validation_curve now supports metadata routing for the fit method of its estimator and for its underlying CV splitter and scorer. By Stefanie Senger #29329

• Fix Metadata is routed correctly to grouped CV splitters via linear_model.RidgeCV and linear_model.RidgeClassifierCV and UnsetMetadataPassedError is fixed for linear_model.RidgeClassifierCV with default scoring. By Stefanie Senger #29634

• Fix Many method arguments which shouldn’t be included in the routing mechanism are now excluded and the set_{method}_request methods are not generated for them. By Adrin Jalali #29920

Dropping official support for PyPy

Due to limited maintainer resources and small number of users, official PyPy support has been dropped. Some parts of scikit-learn may still work but PyPy is not tested anymore in the scikit-learn Continuous Integration. By Loïc Estève #29128

Dropping support for building with setuptools

From scikit-learn 1.6 onwards, support for building with setuptools has been removed. Meson is the only supported way to build scikit-learn, see Building from source for more details. By Loïc Estève #29400

sklearn.base

• Enhancement Added a function base.is_clusterer which determines whether a given estimator is of category clusterer. By Christian Veenhuis #28936

• API Change Passing a class object to is_classifier, is_regressor, is_transformer, and is_outlier_detector is now deprecated. Pass an instance instead. By Adrin Jalali #30122

sklearn.calibration

• API Change cv="prefit" is deprecated for CalibratedClassifierCV. Use FrozenEstimator instead, as CalibratedClassifierCV(FrozenEstimator(estimator)). By Adrin Jalali #30171

sklearn.cluster

• API Change The copy parameter of cluster.Birch was deprecated in 1.6 and will be removed in 1.8. It has no effect as the estimator does not perform in-place operations on the input data. By Yao Xiao #29124

sklearn.compose

• Enhancement sklearn.compose.ColumnTransformer verbose_feature_names_out now accepts string format or callable to generate feature names. By Marc Bresson #28934

sklearn.covariance

• Efficiency covariance.MinCovDet fitting is now slightly faster. By Antony Lee #29835

sklearn.cross_decomposition

• Fix cross_decomposition.PLSRegression properly raises an error when n_components is larger than n_samples. By Thomas Fan #29710

sklearn.datasets

• Feature datasets.fetch_file allows downloading arbitrary data-file from the web. It handles local caching, integrity checks with SHA256 digests and automatic retries in case of HTTP errors. By Olivier Grisel #29354

sklearn.decomposition

• Enhancement LatentDirichletAllocation now has a normalize parameter in transform and fit_transform methods to control whether the document topic distribution is normalized. By Adrin Jalali #30097

• Fix IncrementalPCA will now only raise a ValueError when the number of samples in the input data to partial_fit is less than the number of components on the first call to partial_fit. Subsequent calls to partial_fit no longer face this restriction. By Thomas Gessey-Jones #30224

sklearn.discriminant_analysis

• Fix discriminant_analysis.QuadraticDiscriminantAnalysis will now cause LinAlgWarning in case of collinear variables. These errors can be silenced using the reg_param attribute. By Alihan Zihna #19731

sklearn.ensemble

• Feature ensemble.ExtraTreesClassifier and ensemble.ExtraTreesRegressor now support missing-values in the data matrix X. Missing-values are handled by randomly moving all of the samples to the left, or right child node as the tree is traversed. By Adam Li #28268

• Efficiency Small runtime improvement of fitting ensemble.HistGradientBoostingClassifier and ensemble.HistGradientBoostingRegressor by parallelizing the initial search for bin thresholds. By Christian Lorentzen #28064

• Efficiency ensemble.IsolationForest now runs parallel jobs during predict offering a speedup of up to 2-4x on sample sizes larger than 2000 using joblib. By Adam Li and Sérgio Pereira #28622

• Enhancement The verbosity of ensemble.HistGradientBoostingClassifier and ensemble.HistGradientBoostingRegressor got a more granular control. Now, verbose = 1 prints only summary messages, verbose >= 2 prints the full information as before. By Christian Lorentzen #28179

• API Change The parameter algorithm of ensemble.AdaBoostClassifier is deprecated and will be removed in 1.8. By Jérémie du Boisberranger #29997

sklearn.feature_extraction

• Fix feature_extraction.text.TfidfVectorizer now correctly preserves the dtype of idf_ based on the input data. By Guillaume Lemaitre #30022

sklearn.frozen

• Major Feature FrozenEstimator is now introduced which allows freezing an estimator. This means calling .fit on it has no effect, and doing a clone(frozenestimator) returns the same estimator instead of an unfitted clone. #29705 By Adrin Jalali #29705

sklearn.impute

• Fix impute.KNNImputer excludes samples with nan distances when computing the mean value for uniform weights. By Xuefeng Xu #29135

• Fix When min_value and max_value are array-like and some features are dropped due to keep_empty_features=False, impute.IterativeImputer no longer raises an error and now indexes correctly. By Guntitat Sawadwuthikul #29451

• Fix Fixed impute.IterativeImputer to make sure that it does not skip the iterative process when keep_empty_features is set to True. By Arif Qodari #29779

• API Change Add a warning in impute.SimpleImputer when keep_empty_feature=False and strategy="constant". In this case empty features are not dropped and this behaviour will change in 1.8. By Arthur Courselle and Simon Riou #29950

sklearn.inspection

• Enhancement Add custom_values parameter in inspection.partial_dependence. It enables users to pass their own grid of values at which the partial dependence should be calculated. By Freddy A. Boulton and Stephen Pardy #26202

sklearn.linear_model

• Enhancement The solver="newton-cholesky" in linear_model.LogisticRegression and linear_model.LogisticRegressionCV is extended to support the full multinomial loss in a multiclass setting. By Christian Lorentzen #28840

• Fix In linear_model.Ridge and linear_model.RidgeCV, after fit, the coef_ attribute is now of shape (n_samples,) like other linear models. By Maxwell Liu, Guillaume Lemaitre, and Adrin Jalali #19746

• Fix linear_model.LogisticRegressionCV corrects sample weight handling for the calculation of test scores. By Shruti Nath #29419

• Fix linear_model.LassoCV and linear_model.ElasticNetCV now take sample weights into accounts to define the search grid for the internally tuned alpha hyper-parameter. By John Hopfensperger and Shruti Nath #29442

• Fix linear_model.LogisticRegression, linear_model.PoissonRegressor, linear_model.GammaRegressor, linear_model.TweedieRegressor now take sample weights into account to decide when to fall back to solver='lbfgs' whenever solver='newton-cholesky' becomes numerically unstable. By Antoine Baker #29818

• Fix linear_model.RidgeCV now properly uses predictions on the same scale as the target seen during fit. These predictions are stored in cv_results_ when scoring != None. Previously, the predictions were rescaled by the square root of the sample weights and offset by the mean of the target, leading to an incorrect estimate of the score. By Guillaume Lemaitre, Jérôme Dockes and Hanmin Qin #29842

• Fix linear_model.RidgeCV now properly supports custom multioutput scorers by letting the scorer manage the multioutput averaging. Previously, the predictions and true targets were both squeezed to a 1D array before computing the error. By Guillaume Lemaitre #29884

• Fix linear_model.LinearRegression now sets the cond parameter when calling the scipy.linalg.lstsq solver on dense input data. This ensures more numerically robust results on rank-deficient data. In particular, it empirically fixes the expected equivalence property between fitting with reweighted or with repeated data points. By Antoine Baker #30040

• Fix linear_model.LogisticRegression and and other linear models that accept solver="newton-cholesky" now report the correct number of iterations when they fall back to the "lbfgs" solver because of a rank deficient Hessian matrix. By Olivier Grisel #30100

• Fix SGDOneClassSVM now correctly inherits from OutlierMixin and the tags are correctly set. By Guillaume Lemaitre #30227

• API Change Deprecates copy_X in linear_model.TheilSenRegressor as the parameter has no effect. copy_X will be removed in 1.8. By Adam Li #29105

sklearn.manifold

• Efficiency manifold.locally_linear_embedding and manifold.LocallyLinearEmbedding now allocate more efficiently the memory of sparse matrices in the Hessian, Modified and LTSA methods. By Giorgio Angelotti #28096

sklearn.metrics

• Efficiency sklearn.metrics.classification_report is now faster by caching classification labels. By Adrin Jalali #29738

• Enhancement metrics.RocCurveDisplay.from_estimator, metrics.RocCurveDisplay.from_predictions, metrics.PrecisionRecallDisplay.from_estimator, and metrics.PrecisionRecallDisplay.from_predictions now accept a new keyword despine to remove the top and right spines of the plot in order to make it clearer. By Yao Xiao #26367

• Enhancement sklearn.metrics.check_scoring now accepts raise_exc to specify whether to raise an exception if a subset of the scorers in multimetric scoring fails or to return an error code. By Stefanie Senger #28992

• Fix metrics.roc_auc_score will now correctly return np.nan and warn user if only one class is present in the labels. By Gleb Levitski and Janez Demšar #27412, #30013

• Fix The functions metrics.mean_squared_log_error and metrics.root_mean_squared_log_error now check whether the inputs are within the correct domain for the function \(y=\log(1+x)\), rather than \(y=\log(x)\). The functions metrics.mean_absolute_error, metrics.mean_absolute_percentage_error, metrics.mean_squared_error and metrics.root_mean_squared_error now explicitly check whether a scalar will be returned when multioutput=uniform_average. By Virgil Chan #29709

• API Change The assert_all_finite parameter of functions metrics.pairwise.check_pairwise_arrays and metrics.pairwise_distances is renamed into ensure_all_finite. force_all_finite will be removed in 1.8. By Jérémie du Boisberranger #29404

• API Change scoring="neg_max_error" should be used instead of scoring="max_error" which is now deprecated. By Farid “Freddie” Taba #29462

• API Change The default value of the response_method parameter of metrics.make_scorer will change from None to "predict" and None will be removed in 1.8. In the mean time, None is equivalent to "predict". By Jérémie du Boisberranger #30001

sklearn.model_selection

• Enhancement GroupKFold now has the ability to shuffle groups into different folds when shuffle=True. By Zachary Vealey #28519

• Enhancement There is no need to call fit on a FixedThresholdClassifier if the underlying estimator is already fitted. By Adrin Jalali #30172

• Fix Improve error message when model_selection.RepeatedStratifiedKFold.split is called without a y argument By Anurag Varma #29402

sklearn.neighbors

• Enhancement neighbors.NearestNeighbors, neighbors.KNeighborsClassifier, neighbors.KNeighborsRegressor, neighbors.RadiusNeighborsClassifier, neighbors.RadiusNeighborsRegressor, neighbors.KNeighborsTransformer, neighbors.RadiusNeighborsTransformer, and neighbors.LocalOutlierFactor now work with metric="nan_euclidean", supporting nan inputs. By Carlo Lemos, Guillaume Lemaitre, and Adrin Jalali #25330

• Enhancement Add neighbors.NearestCentroid.decision_function, neighbors.NearestCentroid.predict_proba and neighbors.NearestCentroid.predict_log_proba to the neighbors.NearestCentroid estimator class. Support the case when X is sparse and shrinking_threshold is not None in neighbors.NearestCentroid. By Matthew Ning #26689

• Enhancement Make predict, predict_proba, and score of neighbors.KNeighborsClassifier and neighbors.RadiusNeighborsClassifier accept X=None as input. In this case predictions for all training set points are returned, and points are not included into their own neighbors. By Dmitry Kobak #30047

• Fix neighbors.LocalOutlierFactor raises a warning in the fit method when duplicate values in the training data lead to inaccurate outlier detection. By Henrique Caroço #28773

sklearn.neural_network

• Fix neural_network.MLPRegressor does no longer crash when the model diverges and that early_stopping is enabled. By Marc Bresson #29773

sklearn.pipeline

• Major Feature pipeline.Pipeline can now transform metadata up to the step requiring the metadata, which can be set using the transform_input parameter. By Adrin Jalali #28901

• Enhancement pipeline.Pipeline now warns about not being fitted before calling methods that require the pipeline to be fitted. This warning will become an error in 1.8. By Adrin Jalali #29868

• Fix Fixed an issue with tags and estimator type of Pipeline when pipeline is empty. This allows the HTML representation of an empty pipeline to be rendered correctly. By Gennaro Daniele Acciaro #30203

sklearn.preprocessing

• Enhancement Added warn option to handle_unknown parameter in preprocessing.OneHotEncoder. By Gleb Levitski #28637

• Enhancement The HTML representation of preprocessing.FunctionTransformer will show the function name in the label. By Yao Xiao #29158

• Fix preprocessing.PowerTransformer now uses scipy.special.inv_boxcox to output nan if the input of BoxCox’s inverse is invalid. By Xuefeng Xu #27875

sklearn.semi_supervised

• API Change semi_supervised.SelfTrainingClassifier deprecated the base_estimator parameter in favor of estimator. By Adam Li #28494

sklearn.tree

• Feature tree.ExtraTreeClassifier and tree.ExtraTreeRegressor now support missing-values in the data matrix X. Missing-values are handled by randomly moving all of the samples to the left, or right child node as the tree is traversed. By Adam Li and Loïc Estève #27966, #30318

• Fix Escape double quotes for labels and feature names when exporting trees to Graphviz format. By Santiago M. Mola. #17575

sklearn.utils

• Enhancement utils.check_array now accepts ensure_non_negative to check for negative values in the passed array, until now only available through calling utils.check_non_negative. By Tamara Atanasoska #29540

• Enhancement check_estimator and parametrize_with_checks now check and fail if the classifier has the tags.classifier_tags.multi_class = False tag but does not fail on multi-class data. By Adrin Jalali #29874

• Enhancement utils.validation.check_is_fitted now passes on stateless estimators. An estimator can indicate it’s stateless by setting the requires_fit tag. See Estimator Tags for more information. By Adrin Jalali #29880

• Enhancement Changes to check_estimator and parametrize_with_checks.

  • check_estimator introduces new arguments: on_skip, on_fail, and callback to control the behavior of the check runner. Refer to the API documentation for more details.

  • generate_only=True is deprecated in check_estimator. Use estimator_checks_generator instead.

  • The _xfail_checks estimator tag is now removed, and now in order to indicate which tests are expected to fail, you can pass a dictionary to the check_estimator as the expected_failed_checks parameter. Similarly, the expected_failed_checks parameter in parametrize_with_checks can be used, which is a callable returning a dictionary of the form:

    {
        "check_name": "reason to mark this check as xfail",
    }
    

  By Adrin Jalali #30149

• Fix utils.estimator_checks.parametrize_with_checks and utils.estimator_checks.check_estimator now support estimators that have set_output called on them. By Adrin Jalali #29869

• API Change The assert_all_finite parameter of functions utils.check_array, utils.check_X_y, utils.as_float_array is renamed into ensure_all_finite. force_all_finite will be removed in 1.8. By Jérémie du Boisberranger #29404

• API Change utils.estimator_checks.check_sample_weights_invariance replaced by utils.estimator_checks.check_sample_weight_equivalence_on_dense_data which uses integer (including zero) weights and utils.estimator_checks.check_sample_weight_equivalence_on_sparse_data which does the same on sparse data. By Antoine Baker #29818, #30137

• API Change Using _estimator_type to set the estimator type is deprecated. Inherit from ClassifierMixin, RegressorMixin, TransformerMixin, or OutlierMixin instead. Alternatively, you can set estimator_type in Tags in the __sklearn_tags__ method. By Adrin Jalali #30122

Code and documentation contributors

Thanks to everyone who has contributed to the maintenance and improvement of the project since version 1.7, including:

TODO: update at the time of the release.