Version 1.3.0¶

In Development

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 documentated – 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.

Changed models¶

The following estimators and functions, when fit with the same data and parameters, may produce different models from the previous version. This often occurs due to changes in the modelling logic (bug fixes or enhancements), or in random sampling procedures.

Enhancement multiclass.OutputCodeClassifier.predict now uses a more efficient pairwise distance reduction. As a consequence, the tie-breaking strategy is different and thus the predicted labels may be different. #25196 by Guillaume Lemaitre.
Enhancement The fit_transform method of decomposition.DictionaryLearning is more efficient but may produce different results as in previous versions when transform_algorithm is not the same as fit_algorithm and the number of iterations is small. #24871 by Omar Salman.
Fix Treat more consistently small values in the W and H matrices during the fit and transform steps of decomposition.NMF and decomposition.MiniBatchNMF which can produce different results than previous versions. #25438 by Yotam Avidar-Constantini.
Enhancement The sample_weight parameter now will be used in centroids initialization for cluster.KMeans, cluster.BisectingKMeans and cluster.MiniBatchKMeans. This change will break backward compatibility, since numbers generated from same random seeds will be different. #25752 by Gleb Levitski, Jérémie du Boisberranger, Guillaume Lemaitre.
Fix decomposition.KernelPCA may produce different results through inverse_transform if gamma is None. Now it will be chosen correctly as 1/n_features of the data that it is fitted on, while previously it might be incorrectly chosen as 1/n_features of the data passed to inverse_transform. A new attribute gamma_ is provided for revealing the actual value of gamma used each time the kernel is called. #26337 by Yao Xiao.

Changes impacting all modules¶

Enhancement The get_feature_names_out method of the following classes now raises a NotFittedError if the instance is not fitted. This ensures the error is consistent in all estimators with the get_feature_names_out method.
The NotFittedError displays an informative message asking to fit the instance with the appropriate arguments.

#25294, #25308, #25291, #25367, #25402, by John Pangas, Rahil Parikh , and Alex Buzenet.
Enhancement Added a multi-threaded Cython routine to the compute squared Euclidean distances (sometimes followed by a fused reduction operation) for a pair of datasets consisting of a sparse CSR matrix and a dense NumPy.

This can improve the performance of following functions and estimators:
A typical example of this performance improvement happens when passing a sparse CSR matrix to the predict or transform method of estimators that rely on a dense NumPy representation to store their fitted parameters (or the reverse).

For instance, sklearn.NearestNeighbors.kneighbors is now up to 2 times faster for this case on commonly available laptops.

#25044 by Julien Jerphanion.
Enhancement All estimators that internally rely on OpenMP multi-threading (via Cython) now use a number of threads equal to the number of physical (instead of logical) cores by default. In the past, we observed that using as many threads as logical cores on SMT hosts could sometimes cause severe performance problems depending on the algorithms and the shape of the data. Note that it is still possible to manually adjust the number of threads used by OpenMP as documented in Parallelism.

#26082 by Jérémie du Boisberranger and Olivier Grisel.

Experimental / Under Development¶

Major Feature Metadata routing’s related base methods are included in this release. This feature is only available via the enable_metadata_routing feature flag which can be enabled using sklearn.set_config and sklearn.config_context. For now this feature is mostly useful for third party developers to prepare their code base for metadata routing, and we strongly recommend that they also hide it behind the same feature flag, rather than having it enabled by default. #24027 by Adrin Jalali, Benjamin Bossan, and Omar Salman.

Changelog¶

`sklearn`¶

Feature Added a new option skip_parameter_validation, to the function sklearn.set_config and context manager sklearn.config_context, that allows to skip the validation of the parameters passed to the estimators and public functions. This can be useful to speed up the code but should be used with care because it can lead to unexpected behaviors or raise obscure error messages when setting invalid parameters. #25815 by Jérémie du Boisberranger.

`sklearn.base`¶

Feature A __sklearn_clone__ protocol is now available to override the default behavior of base.clone. #24568 by Thomas Fan.
Fix base.TransformerMixin now currently keeps a namedtuple’s class if transform returns a namedtuple. #26121 by Thomas Fan.

`sklearn.calibration`¶

Fix calibration.CalibratedClassifierCV now does not enforce sample alignment on fit_params. #25805 by Adrin Jalali.

`sklearn.cluster`¶

API Change The sample_weight parameter in predict for cluster.KMeans.predict and cluster.MiniBatchKMeans.predict is now deprecated and will be removed in v1.5. #25251 by Gleb Levitski.
Enhancement The sample_weight parameter now will be used in centroids initialization for cluster.KMeans, cluster.BisectingKMeans and cluster.MiniBatchKMeans. This change will break backward compatibility, since numbers generated from same random seeds will be different. #25752 by Gleb Levitski, Jérémie du Boisberranger, Guillaume Lemaitre.
Major Feature Added cluster.HDBSCAN, a modern hierarchical density-based clustering algorithm. Similarly to cluster.OPTICS, it can be seen as a generalization of DBSCAN by allowing for hierarchical instead of flat clustering, however it varies in its approach from cluster.OPTICS. This algorithm is very robust with respect to its hyperparameters’ values and can be used on a wide variety of data without much, if any, tuning.

This implementation is an adaptation from the original implementation of HDBSCAN in scikit-learn-contrib/hdbscan, by Leland McInnes et al.

#26385 by Meekail Zain

`sklearn.compose`¶

Fix compose.ColumnTransformer raises an informative error when the individual transformers of ColumnTransformer output pandas dataframes with indexes that are not consistent with each other and the output is configured to be pandas. #26286 by Thomas Fan.
Fix compose.ColumnTransformer correctly sets the output of the remainder when set_output is called. #26323 by Thomas Fan.

`sklearn.covariance`¶

API Change Deprecates cov_init in covariance.graphical_lasso in 1.3 since the parameter has no effect. It will be removed in 1.5. #26033 by Genesis Valencia.
API Change Adds costs_ fitted attribute in covariance.GraphicalLasso and covariance.GraphicalLassoCV. #26033 by Genesis Valencia.
API Change Adds covariance parameter in covariance.GraphicalLasso. #26033 by Genesis Valencia.
API Change Adds eps parameter in covariance.GraphicalLasso, covariance.graphical_lasso_path, and covariance.GraphicalLassoCV. #26033 by Genesis Valencia.
Fix Allows alpha=0 in covariance.GraphicalLasso to be consistent with covariance.graphical_lasso. #26033 by Genesis Valencia.
Fix covariance.empirical_covariance now gives an informative error message when input is not appropriate. #26108 by Quentin Barthélemy.

`sklearn.datasets`¶

API Change The data_transposed argument of datasets.make_sparse_coded_signal is deprecated and will be removed in v1.5. #25784 by @Jérémie du Boisberranger.
Fix datasets.fetch_openml returns improved data types when as_frame=True and parser="liac-arff". #26386 by Thomas Fan.
Fix Following the ARFF specs, only the marker "?" is now considered as a missing values when opening ARFF files fetched using datasets.fetch_openml when using the pandas parser. The parameter read_csv_kwargs allows to overwrite this behaviour. #26551 by Guillaume Lemaitre.
Enhancement Allows to overwrite the parameters used to open the ARFF file using the parameter read_csv_kwargs in datasets.fetch_openml when using the pandas parser. #26433 by Guillaume Lemaitre.

`sklearn.decomposition`¶

Enhancement decomposition.DictionaryLearning now accepts the parameter callback for consistency with the function decomposition.dict_learning. #24871 by Omar Salman.
Efficiency decomposition.MiniBatchDictionaryLearning and decomposition.MiniBatchSparsePCA are now faster for small batch sizes by avoiding duplicate validations. #25490 by Jérémie du Boisberranger.
Fix Treat more consistently small values in the W and H matrices during the fit and transform steps of decomposition.NMF and decomposition.MiniBatchNMF which can produce different results than previous versions. #25438 by Yotam Avidar-Constantini.

`sklearn.discriminant_analysis`¶

Enhancement discriminant_analysis.LinearDiscriminantAnalysis now supports the PyTorch. See Array API support (experimental) for more details. #25956 by Thomas Fan.

`sklearn.ensemble`¶

Feature ensemble.HistGradientBoostingRegressor now supports the Gamma deviance loss via loss="gamma". Using the Gamma deviance as loss function comes in handy for modelling skewed distributed, strictly positive valued targets. #22409 by Christian Lorentzen.
Feature Compute a custom out-of-bag score by passing a callable to ensemble.RandomForestClassifier, ensemble.RandomForestRegressor, ensemble.ExtraTreesClassifier and ensemble.ExtraTreesRegressor. #25177 by Tim Head.
Feature ensemble.GradientBoostingClassifier now exposes out-of-bag scores via the oob_scores_ or oob_score_ attributes. #24882 by Ashwin Mathur.
Efficiency ensemble.IsolationForest predict time is now faster (typically by a factor of 8 or more). Internally, the estimator now precomputes decision path lengths per tree at fit time. It is therefore not possible to load an estimator trained with scikit-learn 1.2 to make it predict with scikit-learn 1.3: retraining with scikit-learn 1.3 is required. #25186 by Felipe Breve Siola.
Efficiency ensemble.RandomForestClassifier and ensemble.RandomForestRegressor with warm_start=True now only recomputes out-of-bag scores when there are actually more n_estimators in subsequent fit calls. #26318 by Joshua Choo Yun Keat.
Enhancement ensemble.BaggingClassifier and ensemble.BaggingRegressor expose the allow_nan tag from the underlying estimator. #25506 by Thomas Fan.
Fix ensemble.RandomForestClassifier.fit sets max_samples = 1 when max_samples is a float and round(n_samples * max_samples) < 1. #25601 by Jan Fidor.
Fix ensemble.IsolationForest.fit no longer warns about missing feature names when called with contamination not "auto" on a pandas dataframe. #25931 by Yao Xiao.
Fix ensemble.HistGradientBoostingRegressor and ensemble.HistGradientBoostingClassifier treats negative values for categorical features consistently as missing values, following LightGBM’s and pandas’ conventions. #25629 by Thomas Fan.
Fix Fix deprecation of base_estimator in ensemble.AdaBoostClassifier and ensemble.AdaBoostRegressor that was introduced in #23819. #26242 by Marko Toplak.

`sklearn.exception`¶

Feature Added exception.InconsistentVersionWarning which is raised when a scikit-learn estimator is unpickled with a scikit-learn version that is inconsistent with the sckit-learn version the estimator was pickled with. #25297 by Thomas Fan.

`sklearn.feature_extraction`¶

API Change feature_extraction.image.PatchExtractor now follows the transformer API of scikit-learn. This class is defined as a stateless transformer meaning that it is note required to call fit before calling transform. Parameter validation only happens at fit time. #24230 by Guillaume Lemaitre.

`sklearn.feature_selection`¶

Enhancement All selectors in sklearn.feature_selection will preserve a DataFrame’s dtype when transformed. #25102 by Thomas Fan.
Fix feature_selection.SequentialFeatureSelector’s cv parameter now supports generators. #25973 by Yao Xiao <Charlie-XIAO>.

`sklearn.impute`¶

Enhancement Added the parameter fill_value to impute.IterativeImputer. #25232 by Thijs van Weezel.

`sklearn.inspection`¶

Enhancement Added support for sample_weight in inspection.partial_dependence. This allows for weighted averaging when aggregating for each value of the grid we are making the inspection on. The option is only available when method is set to brute. #25209 by Carlo Lemos.
API Change inspection.partial_dependence returns a utils.Bunch with new key: grid_values. The values key is deprecated in favor of grid_values and the values key will be removed in 1.5. #21809 and #25732 by Thomas Fan.

`sklearn.linear_model`¶

Efficiency Avoid data scaling when sample_weight=None and other unnecessary data copies and unexpected dense to sparse data conversion in linear_model.LinearRegression. #26207 by Olivier Grisel.
Enhancement linear_model.SGDClassifier, linear_model.SGDRegressor and linear_model.SGDOneClassSVM now preserve dtype for numpy.float32. #25587 by Omar Salman.
API Change Deprecates n_iter in favor of max_iter in linear_model.BayesianRidge and linear_model.ARDRegression. n_iter will be removed in scikit-learn 1.5. This change makes those estimators consistent with the rest of estimators. #25697 by John Pangas.
Enhancement The n_iter_ attribute has been included in linear_model.ARDRegression to expose the actual number of iterations required to reach the stopping criterion. #25697 by John Pangas.
Fix Use a more robust criterion to detect convergence of linear_model.LogisticRegression(penalty="l1", solver="liblinear") on linearly separable problems. #25214 by Tom Dupre la Tour.

`sklearn.metrics`¶

Efficiency The computation of the expected mutual information in metrics.adjusted_mutual_info_score is now faster when the number of unique labels is large and its memory usage is reduced in general. #25713 by Kshitij Mathur, Guillaume Lemaitre, Omar Salman and Jérémie du Boisberranger.
Feature Adds zero_division=np.nan to multiple classification metrics: precision_score, recall_score, f1_score, fbeta_score, precision_recall_fscore_support, classification_report. When zero_division=np.nan and there is a zero division, the metric is undefined and is excluded from averaging. When not used for averages, the value returned is np.nan. #25531 by Marc Torrellas Socastro.
Fix metric.manhattan_distances now supports readonly sparse datasets. #25432 by Julien Jerphanion.
Fix Fixed classification_report so that empty input will return np.nan. Previously, “macro avg” and weighted avg would return e.g. f1-score=np.nan and f1-score=0.0, being inconsistent. Now, they both return np.nan. #25531 by Marc Torrellas Socastro.
Fix metric.ndcg_score now gives a meaningful error message for input of length 1. #25672 by Lene Preuss and Wei-Chun Chu.
Enhancement metrics.silhouette_samples nows accepts a sparse matrix of pairwise distances between samples, or a feature array. #18723 by Sahil Gupta and #24677 by Ashwin Mathur.
Enhancement A new parameter drop_intermediate was added to metrics.precision_recall_curve, metrics.PrecisionRecallDisplay.from_estimator, metrics.PrecisionRecallDisplay.from_predictions, which drops some suboptimal thresholds to create lighter precision-recall curves. #24668 by @dberenbaum.
Enhancement metrics.RocCurveDisplay.from_estimator and metrics.RocCurveDisplay.from_predictions now accept two new keywords, plot_chance_level and chance_level_kw to plot the baseline chance level. This line is exposed in the chance_level_ attribute. #25987 by Yao Xiao.
Enhancement metrics.PrecisionRecallDisplay.from_estimator and metrics.PrecisionRecallDisplay.from_predictions now accept two new keywords, plot_chance_level and chance_level_kw to plot the baseline chance level. This line is exposed in the chance_level_ attribute. #26019 by Yao Xiao.
Fix log_loss raises a warning if the values of the parameter y_pred are not normalized, instead of actually normalizing them in the metric. Starting from 1.5 this will raise an error. #25299 by @Omar Salman <OmarManzoor.
API Change The eps parameter of the log_loss has been deprecated and will be removed in 1.5. #25299 by Omar Salman.
Feature metrics.average_precision_score now supports the multiclass case. #17388 by Geoffrey Bolmier and #24769 by Ashwin Mathur.
Fix In metrics.roc_curve, use the threshold value np.inf instead of arbitrary max(y_score) + 1. This threshold is associated with the ROC curve point tpr=0 and fpr=0. #26194 by Guillaume Lemaitre.
Fix The 'matching' metric has been removed when using SciPy>=1.9 to be consistent with scipy.spatial.distance which does not support 'matching' anymore. #26264 by Barata T. Onggo

`sklearn.gaussian_process`¶

Fix gaussian_process.GaussianProcessRegressor has a new argument n_targets, which is used to decide the number of outputs when sampling from the prior distributions. #23099 by Zhehao Liu.

`sklearn.model_selection`¶

Enhancement model_selection.cross_validate accepts a new parameter return_indices to return the train-test indices of each cv split. #25659 by Guillaume Lemaitre.

`sklearn.multioutput`¶

Fix getattr on multioutput.MultiOutputRegressor.partial_fit and multioutput.MultiOutputClassifier.partial_fit now correctly raise an AttributeError if done before calling fit. #26333 by Adrin Jalali.

`sklearn.naive_bayes`¶

Fix naive_bayes.GaussianNB does not raise anymore a ZeroDivisionError when the provided sample_weight reduces the problem to a single class in fit. #24140 by Jonathan Ohayon and Chiara Marmo.

`sklearn.neighbors`¶

Fix Remove support for KulsinskiDistance in neighbors.BallTree. This dissimilarity is not a metric and cannot be supported by the BallTree. #25417 by Guillaume Lemaitre.
Enhancement The performance of neighbors.KNeighborsClassifier.predict and of neighbors.KNeighborsClassifier.predict_proba has been improved when n_neighbors is large and algorithm="brute" with non Euclidean metrics. #24076 by Meekail Zain, Julien Jerphanion.
API Change The support for metrics other than euclidean and manhattan and for callables in neighbors.NearestNeighbors is deprecated and will be removed in version 1.5. #24083 by Valentin Laurent.

`sklearn.neural_network`¶

Fix neural_network.MLPRegressor and neural_network.MLPClassifier reports the right n_iter_ when warm_start=True. It corresponds to the number of iterations performed on the current call to fit instead of the total number of iterations performed since the initialization of the estimator. #25443 by Marvin Krawutschke.

`sklearn.pipeline`¶

Feature pipeline.FeatureUnion can now use indexing notation (e.g. feature_union["scalar"]) to access transformers by name. #25093 by Thomas Fan.
Feature pipeline.FeatureUnion can now access the feature_names_in_ attribute if the X value seen during .fit has a columns attribute and all columns are strings. e.g. when X is a pandas.DataFrame #25220 by Ian Thompson.
Fix pipeline.Pipeline.fit_transform now raises an AttributeError if the last step of the pipeline does not support fit_transform. #26325 by Adrin Jalali.

`sklearn.preprocessing`¶

Major Feature Introduces preprocessing.TargetEncoder which is a categorical encoding based on target mean conditioned on the value of the category. #25334 by Thomas Fan.
Enhancement A new parameter sparse_output was added to SplineTransformer, available as of SciPy 1.8. If sparse_output=True, SplineTransformer returns a sparse CSR matrix. #24145 by Christian Lorentzen.
Enhancement Adds a feature_name_combiner parameter to preprocessing.OneHotEncoder. This specifies a custom callable to create feature names to be returned by get_feature_names_out. The callable combines input arguments (input_feature, category) to a string. #22506 by Mario Kostelac.
Enhancement Added support for sample_weight in preprocessing.KBinsDiscretizer. This allows specifying the parameter sample_weight for each sample to be used while fitting. The option is only available when strategy is set to quantile and kmeans. #24935 by Seladus, Guillaume Lemaitre, and Dea María Léon, #25257 by Gleb Levitski.
Feature preprocessing.OrdinalEncoder now supports grouping infrequent categories into a single feature. Grouping infrequent categories is enabled by specifying how to select infrequent categories with min_frequency or max_categories. #25677 by Thomas Fan.
Enhancement Subsampling through the subsample parameter can now be used in preprocessing.KBinsDiscretizer regardless of the strategy used. #26424 by Jérémie du Boisberranger.
API Change The default value of the subsample parameter of preprocessing.KBinsDiscretizer will change from None to 200_000 in version 1.5 when strategy="kmeans" or strategy="uniform". #26424 by Jérémie du Boisberranger.
Fix AdditiveChi2Sampler is now stateless. The sample_interval_ attribute is deprecated and will be removed in 1.5. #25190 by Vincent Maladière.

`sklearn.svm`¶

API Change dual parameter now accepts auto option for svm.LinearSVC and svm.LinearSVR. #26093 by Gleb Levitski.

`sklearn.tree`¶

Major Feature tree.DecisionTreeRegressor and tree.DecisionTreeClassifier support missing values when splitter='best' and criterion is gini, entropy, or log_loss, for classification or squared_error, friedman_mse, or poisson for regression. #23595, #26376 by Thomas Fan.
Enhancement Adds a class_names parameter to tree.export_text. This allows specifying the parameter class_names for each target class in ascending numerical order. #25387 by William M and crispinlogan.
Fix tree.export_graphviz and tree.export_text now accepts feature_names and class_names as array-like rather than lists. #26289 by Yao Xiao

`sklearn.utils`¶

API Change estimator_checks.check_transformers_unfitted_stateless has been introduced to ensure stateless transformers don’t raise NotFittedError during transform with no prior call to fit or fit_transform. #25190 by Vincent Maladière.
Enhancement preprocessing.PolynomialFeatures now calculates the number of expanded terms a-priori when dealing with sparse csr matrices in order to optimize the choice of dtype for indices and indptr. It can now output csr matrices with np.int32 indices/indptr components when there are few enough elements, and will automatically use np.int64 for sufficiently large matrices. #20524 by niuk-a and #23731 by Meekail Zain
API Change A FutureWarning is now raised when instantiating a class which inherits from a deprecated base class (i.e. decorated by utils.deprecated) and which overrides the __init__ method. #25733 by Brigitta Sipőcz and Jérémie du Boisberranger.
Fix Fixes utils.validation.check_array to properly convert pandas extension arrays. #25813 and #26106 by Thomas Fan.
Fix utils.validation.check_array now supports pandas DataFrames with extension arrays and object dtypes by return an ndarray with object dtype. #25814 by Thomas Fan.

`sklearn.semi_supervised`¶

Enhancement LabelSpreading.fit and LabelPropagation.fit now accepts sparse metrics. #19664 by Kaushik Amar Das.

Miscellaneous¶

Enhancement Replace obsolete exceptions EnvironmentError, IOError and WindowsError. #26466 by Dimitri Papadopoulos ORfanos.

Code and Documentation Contributors¶

Thanks to everyone who has contributed to the maintenance and improvement of the project since version 1.2, including:

TODO: update at the time of the release.