Version 1.4#

For a short description of the main highlights of the release, please refer to Release Highlights for scikit-learn 1.4.

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 1.4.2#

April 2024

This release only includes support for numpy 2.

Version 1.4.1#

February 2024

Metadata Routing#

Fix Fix routing issue with ColumnTransformer when used inside another meta-estimator. #28188 by Adrin Jalali.
Fix No error is raised when no metadata is passed to a metaestimator that includes a sub-estimator which doesn’t support metadata routing. #28256 by Adrin Jalali.
Fix Fix multioutput.MultiOutputRegressor and multioutput.MultiOutputClassifier to work with estimators that don’t consume any metadata when metadata routing is enabled. #28240 by Adrin Jalali.

DataFrame Support#

Enhancement Fix Pandas and Polars dataframe are validated directly without ducktyping checks. #28195 by Thomas Fan.

Changes impacting many modules#

Efficiency Fix Partial revert of #28191 to avoid a performance regression for estimators relying on euclidean pairwise computation with sparse matrices. The impacted estimators are:
#28235 by Julien Jerphanion.
Fix Fixes a bug for all scikit-learn transformers when using set_output with transform set to pandas or polars. The bug could lead to wrong naming of the columns of the returned dataframe. #28262 by Guillaume Lemaitre.
Fix When users try to use a method in StackingClassifier, StackingClassifier, StackingClassifier, SelectFromModel, RFE, SelfTrainingClassifier, OneVsOneClassifier, OutputCodeClassifier or OneVsRestClassifier that their sub-estimators don’t implement, the AttributeError now reraises in the traceback. #28167 by Stefanie Senger.

Changelog#

`sklearn.calibration`#

Fix calibration.CalibratedClassifierCV supports predict_proba with float32 output from the inner estimator. #28247 by Thomas Fan.

`sklearn.cluster`#

Fix cluster.AffinityPropagation now avoids assigning multiple different clusters for equal points. #28121 by Pietro Peterlongo and Yao Xiao.
Fix Avoid infinite loop in cluster.KMeans when the number of clusters is larger than the number of non-duplicate samples. #28165 by Jérémie du Boisberranger.

`sklearn.compose`#

Fix compose.ColumnTransformer now transform into a polars dataframe when verbose_feature_names_out=True and the transformers internally used several times the same columns. Previously, it would raise a due to duplicated column names. #28262 by Guillaume Lemaitre.

`sklearn.ensemble`#

Fix HistGradientBoostingClassifier and HistGradientBoostingRegressor when fitted on pandas DataFrame with extension dtypes, for example pd.Int64Dtype #28385 by Loïc Estève.
Fix Fixes error message raised by ensemble.VotingClassifier when the target is multilabel or multiclass-multioutput in a DataFrame format. #27702 by Guillaume Lemaitre.

`sklearn.impute`#

Fix : impute.SimpleImputer now raises an error in .fit and .transform if fill_value can not be cast to input value dtype with casting='same_kind'. #28365 by Leo Grinsztajn.

`sklearn.inspection`#

Fix inspection.permutation_importance now handles properly sample_weight together with subsampling (i.e. max_features < 1.0). #28184 by Michael Mayer.

`sklearn.linear_model`#

Fix linear_model.ARDRegression now handles pandas input types for predict(X, return_std=True). #28377 by Eddie Bergman.

`sklearn.preprocessing`#

Fix make preprocessing.FunctionTransformer more lenient and overwrite output column names with the get_feature_names_out in the following cases: (i) the input and output column names remain the same (happen when using NumPy ufunc); (ii) the input column names are numbers; (iii) the output will be set to Pandas or Polars dataframe. #28241 by Guillaume Lemaitre.
Fix preprocessing.FunctionTransformer now also warns when set_output is called with transform="polars" and func does not return a Polars dataframe or feature_names_out is not specified. #28263 by Guillaume Lemaitre.
Fix preprocessing.TargetEncoder no longer fails when target_type="continuous" and the input is read-only. In particular, it now works with pandas copy-on-write mode enabled. #28233 by John Hopfensperger.

`sklearn.tree`#

Fix tree.DecisionTreeClassifier and tree.DecisionTreeRegressor are handling missing values properly. The internal criterion was not initialized when no missing values were present in the data, leading to potentially wrong criterion values. #28295 by Guillaume Lemaitre and #28327 by Adam Li.

`sklearn.utils`#

Enhancement Fix utils.metaestimators.available_if now reraises the error from the check function as the cause of the AttributeError. #28198 by Thomas Fan.
Fix utils._safe_indexing now raises a ValueError when X is a Python list and axis=1, as documented in the docstring. #28222 by Guillaume Lemaitre.

Version 1.4.0#

January 2024

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.

Efficiency linear_model.LogisticRegression and linear_model.LogisticRegressionCV now have much better convergence for solvers "lbfgs" and "newton-cg". Both solvers can now reach much higher precision for the coefficients depending on the specified tol. Additionally, lbfgs can make better use of tol, i.e., stop sooner or reach higher precision. Note: The lbfgs is the default solver, so this change might effect many models. This change also means that with this new version of scikit-learn, the resulting coefficients coef_ and intercept_ of your models will change for these two solvers (when fit on the same data again). The amount of change depends on the specified tol, for small values you will get more precise results. #26721 by Christian Lorentzen.
Fix fixes a memory leak seen in PyPy for estimators using the Cython loss functions. #27670 by Guillaume Lemaitre.

Changes impacting all modules#

Major Feature Transformers now support polars output with set_output(transform="polars"). #27315 by Thomas Fan.
Enhancement All estimators now recognizes the column names from any dataframe that adopts the DataFrame Interchange Protocol. Dataframes that return a correct representation through np.asarray(df) is expected to work with our estimators and functions. #26464 by Thomas Fan.
Enhancement The HTML representation of estimators now includes a link to the documentation and is color-coded to denote whether the estimator is fitted or not (unfitted estimators are orange, fitted estimators are blue). #26616 by Riccardo Cappuzzo, Ines Ibnukhsein, Gael Varoquaux, Joel Nothman and Lilian Boulard.
Fix Fixed a bug in most estimators and functions where setting a parameter to a large integer would cause a TypeError. #26648 by Naoise Holohan.

Metadata Routing#

The following models now support metadata routing in one or more or their methods. Refer to the Metadata Routing User Guide for more details.

Feature LarsCV and LassoLarsCV now support metadata routing in their fit method and route metadata to the CV splitter. #27538 by Omar Salman.
Feature multiclass.OneVsRestClassifier, multiclass.OneVsOneClassifier and multiclass.OutputCodeClassifier now support metadata routing in their fit and partial_fit, and route metadata to the underlying estimator’s fit and partial_fit. #27308 by Stefanie Senger.
Feature pipeline.Pipeline now supports metadata routing according to metadata routing user guide. #26789 by Adrin Jalali.
Feature cross_validate, cross_val_score, and cross_val_predict now support metadata routing. The metadata are routed to the estimator’s fit, the scorer, and the CV splitter’s split. The metadata is accepted via the new params parameter. fit_params is deprecated and will be removed in version 1.6. groups parameter is also not accepted as a separate argument when metadata routing is enabled and should be passed via the params parameter. #26896 by Adrin Jalali.
Feature GridSearchCV, RandomizedSearchCV, HalvingGridSearchCV, and HalvingRandomSearchCV now support metadata routing in their fit and score, and route metadata to the underlying estimator’s fit, the CV splitter, and the scorer. #27058 by Adrin Jalali.
Feature ColumnTransformer now supports metadata routing according to metadata routing user guide. #27005 by Adrin Jalali.
Feature linear_model.LogisticRegressionCV now supports metadata routing. linear_model.LogisticRegressionCV.fit now accepts **params which are passed to the underlying splitter and scorer. linear_model.LogisticRegressionCV.score now accepts **score_params which are passed to the underlying scorer. #26525 by Omar Salman.
Feature feature_selection.SelectFromModel now supports metadata routing in fit and partial_fit. #27490 by Stefanie Senger.
Feature linear_model.OrthogonalMatchingPursuitCV now supports metadata routing. Its fit now accepts **fit_params, which are passed to the underlying splitter. #27500 by Stefanie Senger.
Feature ElasticNetCV, LassoCV, MultiTaskElasticNetCV and MultiTaskLassoCV now support metadata routing and route metadata to the CV splitter. #27478 by Omar Salman.
Fix All meta-estimators for which metadata routing is not yet implemented now raise a NotImplementedError on get_metadata_routing and on fit if metadata routing is enabled and any metadata is passed to them. #27389 by Adrin Jalali.

Support for SciPy sparse arrays#

Several estimators are now supporting SciPy sparse arrays. The following functions and classes are impacted:

Functions:

Classes:

cluster.HDBSCAN in #27250 by Yao Xiao;
cluster.KMeans in #27179 by Nurseit Kamchyev;
cluster.MiniBatchKMeans in #27179 by Nurseit Kamchyev;
cluster.OPTICS in #27104 by Maren Westermann and in #27250 by Yao Xiao;
cluster.SpectralClustering in #27161 by Bharat Raghunathan;
decomposition.MiniBatchNMF in #27100 by Isaac Virshup;
decomposition.NMF in #27100 by Isaac Virshup;
feature_extraction.text.TfidfTransformer in #27219 by Yao Xiao;
manifold.Isomap in #27250 by Yao Xiao;
manifold.SpectralEmbedding in #27240 by Yao Xiao;
manifold.TSNE in #27250 by Yao Xiao;
impute.SimpleImputer in #27277 by Yao Xiao;
impute.IterativeImputer in #27277 by Yao Xiao;
impute.KNNImputer in #27277 by Yao Xiao;
kernel_approximation.PolynomialCountSketch in #27301 by Lohit SundaramahaLingam;
neural_network.BernoulliRBM in #27252 by Yao Xiao;
preprocessing.PolynomialFeatures in #27166 by Mohit Joshi;
random_projection.GaussianRandomProjection in #27314 by Stefanie Senger;
random_projection.SparseRandomProjection in #27314 by Stefanie Senger.

Support for Array API#

Several estimators and functions support the Array API. Such changes allows for using the estimators and functions with other libraries such as JAX, CuPy, and PyTorch. This therefore enables some GPU-accelerated computations.

See Array API support (experimental) for more details.

Functions:

sklearn.metrics.accuracy_score and sklearn.metrics.zero_one_loss in #27137 by Edoardo Abati;
sklearn.model_selection.train_test_split in #26855 by Tim Head;
is_multilabel in #27601 by Yaroslav Korobko.

Classes:

decomposition.PCA for the full and randomized solvers (with QR power iterations) in #26315, #27098 and #27431 by Mateusz Sokół, Olivier Grisel and Edoardo Abati;
preprocessing.KernelCenterer in #27556 by Edoardo Abati;
preprocessing.MaxAbsScaler in #27110 by Edoardo Abati;
preprocessing.MinMaxScaler in #26243 by Tim Head;
preprocessing.Normalizer in #27558 by Edoardo Abati.

Private Loss Function Module#

Fix The gradient computation of the binomial log loss is now numerically more stable for very large, in absolute value, input (raw predictions). Before, it could result in np.nan. Among the models that profit from this change are ensemble.GradientBoostingClassifier, ensemble.HistGradientBoostingClassifier and linear_model.LogisticRegression. #28048 by Christian Lorentzen.

Changelog#

`sklearn.base`#

Enhancement base.ClusterMixin.fit_predict and base.OutlierMixin.fit_predict now accept **kwargs which are passed to the fit method of the estimator. #26506 by Adrin Jalali.
Enhancement base.TransformerMixin.fit_transform and base.OutlierMixin.fit_predict now raise a warning if transform / predict consume metadata, but no custom fit_transform / fit_predict is defined in the class inheriting from them correspondingly. #26831 by Adrin Jalali.
Enhancement base.clone now supports dict as input and creates a copy. #26786 by Adrin Jalali.
API Change process_routing now has a different signature. The first two (the object and the method) are positional only, and all metadata are passed as keyword arguments. #26909 by Adrin Jalali.

`sklearn.calibration`#

Enhancement The internal objective and gradient of the sigmoid method of calibration.CalibratedClassifierCV have been replaced by the private loss module. #27185 by Omar Salman.

`sklearn.cluster`#

Fix The degree parameter in the cluster.SpectralClustering constructor now accepts real values instead of only integral values in accordance with the degree parameter of the sklearn.metrics.pairwise.polynomial_kernel. #27668 by Nolan McMahon.
Fix Fixes a bug in cluster.OPTICS where the cluster correction based on predecessor was not using the right indexing. It would lead to inconsistent results depedendent on the order of the data. #26459 by Haoying Zhang and Guillaume Lemaitre.
Fix Improve error message when checking the number of connected components in the fit method of cluster.HDBSCAN. #27678 by Ganesh Tata.
Fix Create copy of precomputed sparse matrix within the fit method of cluster.DBSCAN to avoid in-place modification of the sparse matrix. #27651 by Ganesh Tata.
Fix Raises a proper ValueError when metric="precomputed" and requested storing centers via the parameter store_centers. #27898 by Guillaume Lemaitre.
API Change kdtree and balltree values are now deprecated and are renamed as kd_tree and ball_tree respectively for the algorithm parameter of cluster.HDBSCAN ensuring consistency in naming convention. kdtree and balltree values will be removed in 1.6. #26744 by Shreesha Kumar Bhat.
API Change The option metric=None in cluster.AgglomerativeClustering and cluster.FeatureAgglomeration is deprecated in version 1.4 and will be removed in version 1.6. Use the default value instead. #27828 by Guillaume Lemaitre.

`sklearn.compose`#

Major Feature Adds polars input support to compose.ColumnTransformer through the DataFrame Interchange Protocol. The minimum supported version for polars is 0.19.12. #26683 by Thomas Fan.
Fix cluster.spectral_clustering and cluster.SpectralClustering now raise an explicit error message indicating that sparse matrices and arrays with np.int64 indices are not supported. #27240 by Yao Xiao.
API Change outputs that use pandas extension dtypes and contain pd.NA in ColumnTransformer now result in a FutureWarning and will cause a ValueError in version 1.6, unless the output container has been configured as “pandas” with set_output(transform="pandas"). Before, such outputs resulted in numpy arrays of dtype object containing pd.NA which could not be converted to numpy floats and caused errors when passed to other scikit-learn estimators. #27734 by Jérôme Dockès.

`sklearn.covariance`#

Enhancement Allow covariance.shrunk_covariance to process multiple covariance matrices at once by handling nd-arrays. #25275 by Quentin Barthélemy.
API Change Fix ColumnTransformer now replaces "passthrough" with a corresponding FunctionTransformer in the fitted transformers_ attribute. #27204 by Adrin Jalali.

`sklearn.datasets`#

Enhancement datasets.make_sparse_spd_matrix now uses a more memory- efficient sparse layout. It also accepts a new keyword sparse_format that allows specifying the output format of the sparse matrix. By default sparse_format=None, which returns a dense numpy ndarray as before. #27438 by Yao Xiao.
Fix datasets.dump_svmlight_file now does not raise ValueError when X is read-only, e.g., a numpy.memmap instance. #28111 by Yao Xiao.
API Change datasets.make_sparse_spd_matrix deprecated the keyword argument dim in favor of n_dim. dim will be removed in version 1.6. #27718 by Adam Li.

`sklearn.decomposition`#

Feature decomposition.PCA now supports scipy.sparse.sparray and scipy.sparse.spmatrix inputs when using the arpack solver. When used on sparse data like datasets.fetch_20newsgroups_vectorized this can lead to speed-ups of 100x (single threaded) and 70x lower memory usage. Based on Alexander Tarashansky’s implementation in scanpy. #18689 by Isaac Virshup and Andrey Portnoy.
Enhancement An “auto” option was added to the n_components parameter of decomposition.non_negative_factorization, decomposition.NMF and decomposition.MiniBatchNMF to automatically infer the number of components from W or H shapes when using a custom initialization. The default value of this parameter will change from None to auto in version 1.6. #26634 by Alexandre Landeau and Alexandre Vigny.
Fix decomposition.dict_learning_online does not ignore anymore the parameter max_iter. #27834 by Guillaume Lemaitre.
Fix The degree parameter in the decomposition.KernelPCA constructor now accepts real values instead of only integral values in accordance with the degree parameter of the sklearn.metrics.pairwise.polynomial_kernel. #27668 by Nolan McMahon.
API Change The option max_iter=None in decomposition.MiniBatchDictionaryLearning, decomposition.MiniBatchSparsePCA, and decomposition.dict_learning_online is deprecated and will be removed in version 1.6. Use the default value instead. #27834 by Guillaume Lemaitre.

`sklearn.ensemble`#

Major Feature ensemble.RandomForestClassifier and ensemble.RandomForestRegressor support missing values when the criterion is gini, entropy, or log_loss, for classification or squared_error, friedman_mse, or poisson for regression. #26391 by Thomas Fan.
Major Feature ensemble.HistGradientBoostingClassifier and ensemble.HistGradientBoostingRegressor supports categorical_features="from_dtype", which treats columns with Pandas or Polars Categorical dtype as categories in the algorithm. categorical_features="from_dtype" will become the default in v1.6. Categorical features no longer need to be encoded with numbers. When categorical features are numbers, the maximum value no longer needs to be smaller than max_bins; only the number of (unique) categories must be smaller than max_bins. #26411 by Thomas Fan and #27835 by Jérôme Dockès.
Major Feature ensemble.HistGradientBoostingClassifier and ensemble.HistGradientBoostingRegressor got the new parameter max_features to specify the proportion of randomly chosen features considered in each split. #27139 by Christian Lorentzen.
Feature ensemble.RandomForestClassifier, ensemble.RandomForestRegressor, ensemble.ExtraTreesClassifier and ensemble.ExtraTreesRegressor now support monotonic constraints, useful when features are supposed to have a positive/negative effect on the target. Missing values in the train data and multi-output targets are not supported. #13649 by Samuel Ronsin, initiated by Patrick O’Reilly.
Efficiency ensemble.HistGradientBoostingClassifier and ensemble.HistGradientBoostingRegressor are now a bit faster by reusing the parent node’s histogram as children node’s histogram in the subtraction trick. In effect, less memory has to be allocated and deallocated. #27865 by Christian Lorentzen.
Efficiency ensemble.GradientBoostingClassifier is faster, for binary and in particular for multiclass problems thanks to the private loss function module. #26278 and #28095 by Christian Lorentzen.
Efficiency Improves runtime and memory usage for ensemble.GradientBoostingClassifier and ensemble.GradientBoostingRegressor when trained on sparse data. #26957 by Thomas Fan.
Efficiency ensemble.HistGradientBoostingClassifier and ensemble.HistGradientBoostingRegressor is now faster when scoring is a predefined metric listed in metrics.get_scorer_names and early stopping is enabled. #26163 by Thomas Fan.
Enhancement A fitted property, estimators_samples_, was added to all Forest methods, including ensemble.RandomForestClassifier, ensemble.RandomForestRegressor, ensemble.ExtraTreesClassifier and ensemble.ExtraTreesRegressor, which allows to retrieve the training sample indices used for each tree estimator. #26736 by Adam Li.
Fix Fixes ensemble.IsolationForest when the input is a sparse matrix and contamination is set to a float value. #27645 by Guillaume Lemaitre.
Fix Raises a ValueError in ensemble.RandomForestRegressor and ensemble.ExtraTreesRegressor when requesting OOB score with multioutput model for the targets being all rounded to integer. It was recognized as a multiclass problem. #27817 by Daniele Ongari
Fix Changes estimator tags to acknowledge that ensemble.VotingClassifier, ensemble.VotingRegressor, ensemble.StackingClassifier, ensemble.StackingRegressor, support missing values if all estimators support missing values. #27710 by Guillaume Lemaitre.
Fix Support loading pickles of ensemble.HistGradientBoostingClassifier and ensemble.HistGradientBoostingRegressor when the pickle has been generated on a platform with a different bitness. A typical example is to train and pickle the model on 64 bit machine and load the model on a 32 bit machine for prediction. #28074 by Christian Lorentzen and Loïc Estève.
API Change In ensemble.AdaBoostClassifier, the algorithm argument SAMME.R was deprecated and will be removed in 1.6. #26830 by Stefanie Senger.

`sklearn.feature_extraction`#

API Change Changed error type from AttributeError to exceptions.NotFittedError in unfitted instances of feature_extraction.DictVectorizer for the following methods: feature_extraction.DictVectorizer.inverse_transform, feature_extraction.DictVectorizer.restrict, feature_extraction.DictVectorizer.transform. #24838 by Lorenz Hertel.

`sklearn.feature_selection`#

Enhancement feature_selection.SelectKBest, feature_selection.SelectPercentile, and feature_selection.GenericUnivariateSelect now support unsupervised feature selection by providing a score_func taking X and y=None. #27721 by Guillaume Lemaitre.
Enhancement feature_selection.SelectKBest and feature_selection.GenericUnivariateSelect with mode='k_best' now shows a warning when k is greater than the number of features. #27841 by Thomas Fan.
Fix feature_selection.RFE and feature_selection.RFECV do not check for nans during input validation. #21807 by Thomas Fan.

`sklearn.inspection`#

Enhancement inspection.DecisionBoundaryDisplay now accepts a parameter class_of_interest to select the class of interest when plotting the response provided by response_method="predict_proba" or response_method="decision_function". It allows to plot the decision boundary for both binary and multiclass classifiers. #27291 by Guillaume Lemaitre.
Fix inspection.DecisionBoundaryDisplay.from_estimator and inspection.PartialDependenceDisplay.from_estimator now return the correct type for subclasses. #27675 by John Cant.
API Change inspection.DecisionBoundaryDisplay raise an AttributeError instead of a ValueError when an estimator does not implement the requested response method. #27291 by Guillaume Lemaitre.

`sklearn.kernel_ridge`#

Fix The degree parameter in the kernel_ridge.KernelRidge constructor now accepts real values instead of only integral values in accordance with the degree parameter of the sklearn.metrics.pairwise.polynomial_kernel. #27668 by Nolan McMahon.

`sklearn.linear_model`#

Efficiency linear_model.LogisticRegression and linear_model.LogisticRegressionCV now have much better convergence for solvers "lbfgs" and "newton-cg". Both solvers can now reach much higher precision for the coefficients depending on the specified tol. Additionally, lbfgs can make better use of tol, i.e., stop sooner or reach higher precision. This is accomplished by better scaling of the objective function, i.e., using average per sample losses instead of sum of per sample losses. #26721 by Christian Lorentzen.
Efficiency linear_model.LogisticRegression and linear_model.LogisticRegressionCV with solver "newton-cg" can now be considerably faster for some data and parameter settings. This is accomplished by a better line search convergence check for negligible loss improvements that takes into account gradient information. #26721 by Christian Lorentzen.
Efficiency Solver "newton-cg" in linear_model.LogisticRegression and linear_model.LogisticRegressionCV uses a little less memory. The effect is proportional to the number of coefficients (n_features * n_classes). #27417 by Christian Lorentzen.
Fix Ensure that the sigma_ attribute of linear_model.ARDRegression and linear_model.BayesianRidge always has a float32 dtype when fitted on float32 data, even with the type promotion rules of NumPy 2. #27899 by Olivier Grisel.
API Change The attribute loss_function_ of linear_model.SGDClassifier and linear_model.SGDOneClassSVM has been deprecated and will be removed in version 1.6. #27979 by Christian Lorentzen.

`sklearn.metrics`#

Efficiency Computing pairwise distances via metrics.DistanceMetric for CSR x CSR, Dense x CSR, and CSR x Dense datasets is now 1.5x faster. #26765 by Meekail Zain.
Efficiency Computing distances via metrics.DistanceMetric for CSR x CSR, Dense x CSR, and CSR x Dense now uses ~50% less memory, and outputs distances in the same dtype as the provided data. #27006 by Meekail Zain.
Enhancement Improve the rendering of the plot obtained with the metrics.PrecisionRecallDisplay and metrics.RocCurveDisplay classes. the x- and y-axis limits are set to [0, 1] and the aspect ratio between both axis is set to be 1 to get a square plot. #26366 by Mojdeh Rastgoo.
Enhancement Added neg_root_mean_squared_log_error_scorer as scorer #26734 by Alejandro Martin Gil.
Enhancement metrics.confusion_matrix now warns when only one label was found in y_true and y_pred. #27650 by Lucy Liu.
Fix computing pairwise distances with metrics.pairwise.euclidean_distances no longer raises an exception when X is provided as a float64 array and X_norm_squared as a float32 array. #27624 by Jérôme Dockès.
Fix f1_score now provides correct values when handling various cases in which division by zero occurs by using a formulation that does not depend on the precision and recall values. #27577 by Omar Salman and Guillaume Lemaitre.
Fix metrics.make_scorer now raises an error when using a regressor on a scorer requesting a non-thresholded decision function (from decision_function or predict_proba). Such scorer are specific to classification. #26840 by Guillaume Lemaitre.
Fix metrics.DetCurveDisplay.from_predictions, metrics.PrecisionRecallDisplay.from_predictions, metrics.PredictionErrorDisplay.from_predictions, and metrics.RocCurveDisplay.from_predictions now return the correct type for subclasses. #27675 by John Cant.
API Change Deprecated needs_threshold and needs_proba from metrics.make_scorer. These parameters will be removed in version 1.6. Instead, use response_method that accepts "predict", "predict_proba" or "decision_function" or a list of such values. needs_proba=True is equivalent to response_method="predict_proba" and needs_threshold=True is equivalent to response_method=("decision_function", "predict_proba"). #26840 by Guillaume Lemaitre.
API Change The squared parameter of metrics.mean_squared_error and metrics.mean_squared_log_error is deprecated and will be removed in 1.6. Use the new functions metrics.root_mean_squared_error and metrics.root_mean_squared_log_error instead. #26734 by Alejandro Martin Gil.

`sklearn.model_selection`#

Enhancement model_selection.learning_curve raises a warning when every cross validation fold fails. #26299 by Rahil Parikh.
Fix model_selection.GridSearchCV, model_selection.RandomizedSearchCV, and model_selection.HalvingGridSearchCV now don’t change the given object in the parameter grid if it’s an estimator. #26786 by Adrin Jalali.

`sklearn.multioutput`#

Enhancement Add method predict_log_proba to multioutput.ClassifierChain. #27720 by Guillaume Lemaitre.

`sklearn.neighbors`#

Efficiency sklearn.neighbors.KNeighborsRegressor.predict and sklearn.neighbors.KNeighborsClassifier.predict_proba now efficiently support pairs of dense and sparse datasets. #27018 by Julien Jerphanion.
Efficiency The performance of neighbors.RadiusNeighborsClassifier.predict and of neighbors.RadiusNeighborsClassifier.predict_proba has been improved when radius is large and algorithm="brute" with non-Euclidean metrics. #26828 by Omar Salman.
Fix Improve error message for neighbors.LocalOutlierFactor when it is invoked with n_samples=n_neighbors. #23317 by Bharat Raghunathan.
Fix neighbors.KNeighborsClassifier.predict and neighbors.KNeighborsClassifier.predict_proba now raises an error when the weights of all neighbors of some sample are zero. This can happen when weights is a user-defined function. #26410 by Yao Xiao.
API Change neighbors.KNeighborsRegressor now accepts metrics.DistanceMetric objects directly via the metric keyword argument allowing for the use of accelerated third-party metrics.DistanceMetric objects. #26267 by Meekail Zain.

`sklearn.preprocessing`#

Efficiency preprocessing.OrdinalEncoder avoids calculating missing indices twice to improve efficiency. #27017 by Xuefeng Xu.
Efficiency Improves efficiency in preprocessing.OneHotEncoder and preprocessing.OrdinalEncoder in checking nan. #27760 by Xuefeng Xu.
Enhancement Improves warnings in preprocessing.FunctionTransformer when func returns a pandas dataframe and the output is configured to be pandas. #26944 by Thomas Fan.
Enhancement preprocessing.TargetEncoder now supports target_type ‘multiclass’. #26674 by Lucy Liu.
Fix preprocessing.OneHotEncoder and preprocessing.OrdinalEncoder raise an exception when nan is a category and is not the last in the user’s provided categories. #27309 by Xuefeng Xu.
Fix preprocessing.OneHotEncoder and preprocessing.OrdinalEncoder raise an exception if the user provided categories contain duplicates. #27328 by Xuefeng Xu.
Fix preprocessing.FunctionTransformer raises an error at transform if the output of get_feature_names_out is not consistent with the column names of the output container if those are defined. #27801 by Guillaume Lemaitre.
Fix Raise a NotFittedError in preprocessing.OrdinalEncoder when calling transform without calling fit since categories always requires to be checked. #27821 by Guillaume Lemaitre.

`sklearn.tree`#

Feature tree.DecisionTreeClassifier, tree.DecisionTreeRegressor, tree.ExtraTreeClassifier and tree.ExtraTreeRegressor now support monotonic constraints, useful when features are supposed to have a positive/negative effect on the target. Missing values in the train data and multi-output targets are not supported. #13649 by Samuel Ronsin, initiated by Patrick O’Reilly.

`sklearn.utils`#

Enhancement sklearn.utils.estimator_html_repr dynamically adapts diagram colors based on the browser’s prefers-color-scheme, providing improved adaptability to dark mode environments. #26862 by Andrew Goh Yisheng, Thomas Fan, Adrin Jalali.
Enhancement MetadataRequest and MetadataRouter now have a consumes method which can be used to check whether a given set of parameters would be consumed. #26831 by Adrin Jalali.
Enhancement Make sklearn.utils.check_array attempt to output int32-indexed CSR and COO arrays when converting from DIA arrays if the number of non-zero entries is small enough. This ensures that estimators implemented in Cython and that do not accept int64-indexed sparse datastucture, now consistently accept the same sparse input formats for SciPy sparse matrices and arrays. #27372 by Guillaume Lemaitre.
Fix sklearn.utils.check_array should accept both matrix and array from the sparse SciPy module. The previous implementation would fail if copy=True by calling specific NumPy np.may_share_memory that does not work with SciPy sparse array and does not return the correct result for SciPy sparse matrix. #27336 by Guillaume Lemaitre.
Fix check_estimators_pickle with readonly_memmap=True now relies on joblib’s own capability to allocate aligned memory mapped arrays when loading a serialized estimator instead of calling a dedicated private function that would crash when OpenBLAS misdetects the CPU architecture. #27614 by Olivier Grisel.
Fix Error message in check_array when a sparse matrix was passed but accept_sparse is False now suggests to use .toarray() and not X.toarray(). #27757 by Lucy Liu.
Fix Fix the function check_array to output the right error message when the input is a Series instead of a DataFrame. #28090 by Stan Furrer and Yao Xiao.
API Change sklearn.extmath.log_logistic is deprecated and will be removed in 1.6. Use -np.logaddexp(0, -x) instead. #27544 by Christian Lorentzen.

Code and documentation contributors

Thanks to everyone who has contributed to the maintenance and improvement of the project since version 1.3, including:

101AlexMartin, Abhishek Singh Kushwah, Adam Li, Adarsh Wase, Adrin Jalali, Advik Sinha, Alex, Alexander Al-Feghali, Alexis IMBERT, AlexL, Alex Molas, Anam Fatima, Andrew Goh, andyscanzio, Aniket Patil, Artem Kislovskiy, Arturo Amor, ashah002, avm19, Ben Holmes, Ben Mares, Benoit Chevallier-Mames, Bharat Raghunathan, Binesh Bannerjee, Brendan Lu, Brevin Kunde, Camille Troillard, Carlo Lemos, Chad Parmet, Christian Clauss, Christian Lorentzen, Christian Veenhuis, Christos Aridas, Cindy Liang, Claudio Salvatore Arcidiacono, Connor Boyle, cynthias13w, DaminK, Daniele Ongari, Daniel Schmitz, Daniel Tinoco, David Brochart, Deborah L. Haar, DevanshKyada27, Dimitri Papadopoulos Orfanos, Dmitry Nesterov, DUONG, Edoardo Abati, Eitan Hemed, Elabonga Atuo, Elisabeth Günther, Emma Carballal, Emmanuel Ferdman, epimorphic, Erwan Le Floch, Fabian Egli, Filip Karlo Došilović, Florian Idelberger, Franck Charras, Gael Varoquaux, Ganesh Tata, Gleb Levitski, Guillaume Lemaitre, Haoying Zhang, Harmanan Kohli, Ily, ioangatop, IsaacTrost, Isaac Virshup, Iwona Zdzieblo, Jakub Kaczmarzyk, James McDermott, Jarrod Millman, JB Mountford, Jérémie du Boisberranger, Jérôme Dockès, Jiawei Zhang, Joel Nothman, John Cant, John Hopfensperger, Jona Sassenhagen, Jon Nordby, Julien Jerphanion, Kennedy Waweru, kevin moore, Kian Eliasi, Kishan Ved, Konstantinos Pitas, Koustav Ghosh, Kushan Sharma, ldwy4, Linus, Lohit SundaramahaLingam, Loic Esteve, Lorenz, Louis Fouquet, Lucy Liu, Luis Silvestrin, Lukáš Folwarczný, Lukas Geiger, Malte Londschien, Marcus Fraaß, Marek Hanuš, Maren Westermann, Mark Elliot, Martin Larralde, Mateusz Sokół, mathurinm, mecopur, Meekail Zain, Michael Higgins, Miki Watanabe, Milton Gomez, MN193, Mohammed Hamdy, Mohit Joshi, mrastgoo, Naman Dhingra, Naoise Holohan, Narendra Singh dangi, Noa Malem-Shinitski, Nolan, Nurseit Kamchyev, Oleksii Kachaiev, Olivier Grisel, Omar Salman, partev, Peter Hull, Peter Steinbach, Pierre de Fréminville, Pooja Subramaniam, Puneeth K, qmarcou, Quentin Barthélemy, Rahil Parikh, Rahul Mahajan, Raj Pulapakura, Raphael, Ricardo Peres, Riccardo Cappuzzo, Roman Lutz, Salim Dohri, Samuel O. Ronsin, Sandip Dutta, Sayed Qaiser Ali, scaja, scikit-learn-bot, Sebastian Berg, Shreesha Kumar Bhat, Shubhal Gupta, Søren Fuglede Jørgensen, Stefanie Senger, Tamara, Tanjina Afroj, THARAK HEGDE, thebabush, Thomas J. Fan, Thomas Roehr, Tialo, Tim Head, tongyu, Venkatachalam N, Vijeth Moudgalya, Vincent M, Vivek Reddy P, Vladimir Fokow, Xiao Yuan, Xuefeng Xu, Yang Tao, Yao Xiao, Yuchen Zhou, Yuusuke Hiramatsu