Version 1.8

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 1.8.dev0

October 2025

Changes impacting many modules

• Efficiency Improved CPU and memory usage in estimators and metric functions that rely on weighted percentiles and better match NumPy and Scipy (un-weighted) implementations of percentiles. By Lucy Liu #31775

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 sklearn.preprocessing.StandardScaler now supports Array API compliant inputs. By Alexander Fabisch, Edoardo Abati, Olivier Grisel and Charles Hill. #27113

• Feature metrics.pairwise.pairwise_kernels for any kernel except “laplacian” and metrics.pairwise_distances for metrics “cosine”, “euclidean” and “l2” now support array API inputs. By Emily Chen and Lucy Liu #29822

• Feature sklearn.metrics.confusion_matrix now supports Array API compatible inputs. By Stefanie Senger #30562

• Feature sklearn.mixture.GaussianMixture with init_params="random" or init_params="random_from_data" and warm_start=False now supports Array API compatible inputs. By Stefanie Senger and Loïc Estève #30777

• Feature sklearn.metrics.roc_curve now supports Array API compatible inputs. By Thomas Li #30878

• Feature preprocessing.PolynomialFeatures now supports array API compatible inputs. By Omar Salman #31580

• Feature sklearn.model_selection.cross_val_predict now supports array API compatible inputs. By Omar Salman #32270

Metadata routing

Refer to the Metadata Routing User Guide for more details.

• Fix Fixed an issue where passing sample_weight to a Pipeline inside a GridSearchCV would raise an error with metadata routing enabled. By Adrin Jalali. #31898

sklearn.base

• Feature Refactored dir in BaseEstimator to recognize condition check in available_if. By John Hendricks and Miguel Parece. #31928

sklearn.calibration

• Feature Added temperature scaling method in calibration.CalibratedClassifierCV. By Virgil Chan and Christian Lorentzen. #31068

sklearn.cluster

• Efficiency cluster.kmeans_plusplus now uses np.cumsum directly without extra numerical stability checks and without casting to np.float64. By Tiziano Zito #31991

• Fix The default value of the copy parameter in cluster.HDBSCAN will change from False to True in 1.10 to avoid data modification and maintain consistency with other estimators. By Sarthak Puri. #31973

sklearn.covariance

• Efficiency sklearn.covariance.GraphicalLasso, sklearn.covariance.GraphicalLassoCV and sklearn.covariance.graphical_lasso with mode="cd" profit from the fit time performance improvement of sklearn.linear_model.Lasso by means of gap safe screening rules. By Christian Lorentzen. #31987

• Fix Fixed uncontrollable randomness in sklearn.covariance.GraphicalLasso, sklearn.covariance.GraphicalLassoCV and sklearn.covariance.graphical_lasso. For mode="cd", they now use cyclic coordinate descent. Before, it was random coordinate descent with uncontrollable random number seeding. By Christian Lorentzen. #31987

sklearn.decomposition

• Efficiency sklearn.decomposition.DictionaryLearning and sklearn.decomposition.MiniBatchDictionaryLearning with fit_algorithm="cd", sklearn.decomposition.SparseCoder with transform_algorithm="lasso_cd", sklearn.decomposition.MiniBatchSparsePCA, sklearn.decomposition.SparsePCA, sklearn.decomposition.dict_learning and sklearn.decomposition.dict_learning_online with method="cd", sklearn.decomposition.sparse_encode with algorithm="lasso_cd" all profit from the fit time performance improvement of sklearn.linear_model.Lasso by means of gap safe screening rules. By Christian Lorentzen. #31987

• Enhancement decomposition.SparseCoder now follows the transformer API of scikit-learn. In addition, the fit method now validates the input and parameters. By François Paugam. #32077

• Fix Add input checks to the inverse_transform method of decomposition.PCA and decomposition.IncrementalPCA. #29310 by Ian Faust. #29310

sklearn.ensemble

• Fix ensemble.BaggingClassifier, ensemble.BaggingRegressor and ensemble.IsolationForest now use sample_weight to draw the samples instead of forwarding them multiplied by a uniformly sampled mask to the underlying estimators. Furthermore, max_samples is now interpreted as a fraction of sample_weight.sum() instead of X.shape[0] when passed as a float. By Antoine Baker. #31414

sklearn.gaussian_process

• Efficiency make GaussianProcessRegressor.predict faster when return_cov and return_std are both False. By Rafael Ayllón Gavilán. #31431

sklearn.linear_model

• Efficiency linear_model.ElasticNet and linear_model.Lasso with precompute=False use less memory for dense X and are a bit faster. Previously, they used twice the memory of X even for Fortran-contiguous X. By Christian Lorentzen #31665

• Efficiency linear_model.ElasticNet and linear_model.Lasso avoid double input checking and are therefore a bit faster. By Christian Lorentzen. #31848

• Efficiency linear_model.ElasticNet, linear_model.ElasticNetCV, linear_model.Lasso, linear_model.LassoCV, linear_model.MultiTaskElasticNet, linear_model.MultiTaskElasticNetCV, linear_model.MultiTaskLasso and linear_model.MultiTaskLassoCV are faster to fit by avoiding a BLAS level 1 (axpy) call in the innermost loop. Same for functions linear_model.enet_path and linear_model.lasso_path. By Christian Lorentzen #31956 and #31880

• Efficiency linear_model.ElasticNetCV, linear_model.LassoCV, linear_model.MultiTaskElasticNetCV and linear_model.MultiTaskLassoCV avoid an additional copy of X with default copy_X=True. By Christian Lorentzen. #31946

• Efficiency linear_model.ElasticNet, linear_model.ElasticNetCV, linear_model.Lasso, linear_model.LassoCV, linear_model.MultiTaskElasticNetCV, linear_model.MultiTaskLassoCV as well as linear_model.lasso_path and linear_model.enet_path now implement gap safe screening rules in the coordinate descent solver for dense and sparse X. The speedup of fitting time is particularly pronounced (10-times is possible) when computing regularization paths like the *CV-variants of the above estimators do. There is now an additional check of the stopping criterion before entering the main loop of descent steps. As the stopping criterion requires the computation of the dual gap, the screening happens whenever the dual gap is computed. By Christian Lorentzen #31882, #31986, #31987 and #32014

• Enhancement linear_model.ElasticNet, linear_model.ElasticNetCV, linear_model.Lasso, linear_model.LassoCV, MultiTaskElasticNet, MultiTaskElasticNetCV, MultiTaskLasso, MultiTaskLassoCV, as well as linear_model.enet_path and linear_model.lasso_path now use dual gap <= tol instead of dual gap < tol as stopping criterion. The resulting coefficients might differ to previous versions of scikit-learn in rare cases. By Christian Lorentzen. #31906

• Fix Fix the convergence criteria for SGD models, to avoid premature convergence when tol != None. This primarily impacts SGDOneClassSVM but also affects SGDClassifier and SGDRegressor. Before this fix, only the loss function without penalty was used as the convergence check, whereas now, the full objective with regularization is used. By Guillaume Lemaitre and kostayScr #31856

• Fix The allowed parameter range for the initial learning rate eta0 in linear_model.SGDClassifier, linear_model.SGDOneClassSVM, linear_model.SGDRegressor and linear_model.Perceptron changed from non-negative numbers to strictly positive numbers. As a consequence, the default eta0 of linear_model.SGDClassifier and linear_model.SGDOneClassSVM changed from 0 to 0.01. But note that eta0 is not used by the default learning rate “optimal” of those two estimators. By Christian Lorentzen. #31933

• API Change linear_model.PassiveAggressiveClassifier and linear_model.PassiveAggressiveRegressor are deprecated and will be removed in 1.10. Equivalent estimators are available with linear_model.SGDClassifier and SGDRegressor, both of which expose the options learning_rate="pa1" and "pa2". The parameter eta0 can be used to specify the aggressiveness parameter of the Passive-Aggressive-Algorithms, called C in the reference paper. By Christian Lorentzen #31932 and #29097

• API Change linear_model.SGDClassifier, linear_model.SGDRegressor, and linear_model.SGDOneClassSVM now deprecate negative values for the power_t parameter. Using a negative value will raise a warning in version 1.8 and will raise an error in version 1.10. A value in the range [0.0, inf) must be used instead. By Ritvi Alagusankar #31474

sklearn.manifold

• Major Feature manifold.ClassicalMDS was implemented to perform classical MDS (eigendecomposition of the double-centered distance matrix). By Dmitry Kobak and Meekail Zain #31322

sklearn.metrics

• Feature metrics.d2_brier_score has been added which calculates the D^2 for the Brier score. By Omar Salman. #28971

• Enhancement metrics.median_absolute_error now supports Array API compatible inputs. By Lucy Liu. #31406

• Enhancement Improved the error message for sparse inputs for the following metrics: metrics.accuracy_score, metrics.multilabel_confusion_matrix, metrics.jaccard_score, metrics.zero_one_loss, metrics.f1_score, metrics.fbeta_score, metrics.precision_recall_fscore_support, metrics.class_likelihood_ratios, metrics.precision_score, metrics.recall_score, metrics.classification_report, metrics.hamming_loss. By Lucy Liu. #32047

• Fix metrics.median_absolute_error now uses _averaged_weighted_percentile instead of _weighted_percentile to calculate median when sample_weight is not None. This is equivalent to using the “averaged_inverted_cdf” instead of the “inverted_cdf” quantile method, which gives results equivalent to numpy.median if equal weights used. By Lucy Liu #30787

• Fix Additional sample_weight checking has been added to metrics.accuracy_score, metrics.balanced_accuracy_score, metrics.brier_score_loss, metrics.class_likelihood_ratios, metrics.classification_report, metrics.cohen_kappa_score, metrics.confusion_matrix, metrics.f1_score, metrics.fbeta_score, metrics.hamming_loss, metrics.jaccard_score, metrics.matthews_corrcoef, metrics.multilabel_confusion_matrix, metrics.precision_recall_fscore_support, metrics.precision_score, metrics.recall_score and metrics.zero_one_loss. sample_weight can only be 1D, consistent to y_true and y_pred in length,and all values must be finite and not complex. By Lucy Liu. #31701

• Fix y_pred is deprecated in favour of y_score in metrics.DetCurveDisplay.from_predictions and metrics.PrecisionRecallDisplay.from_predictions. y_pred will be removed in v1.10. By Luis #31764

• Fix repr on a scorer which has been created with a partial score_func now correctly works and uses the repr of the given partial object. By Adrin Jalali. #31891

• API Change metrics.cluster.entropy is deprecated and will be removed in v1.10. By Lucy Liu #31294

• API Change The estimator_name parameter is deprecated in favour of name in metrics.PrecisionRecallDisplay and will be removed in 1.10. By Lucy Liu. #32310

sklearn.multiclass

• Fix Fix tie-breaking behavior in multiclass.OneVsRestClassifier to match np.argmax tie-breaking behavior. By Lakshmi Krishnan. #15504

sklearn.preprocessing

• Enhancement preprocessing.SplineTransformer can now handle missing values with the parameter handle_missing. By Stefanie Senger. #28043

• Enhancement The preprocessing.PowerTransformer now returns a warning when NaN values are encountered in the inverse transform, inverse_transform, typically caused by extremely skewed data. By Roberto Mourao #29307

• Enhancement preprocessing.MaxAbsScaler can now clip out-of-range values in held-out data with the parameter clip. By Hleb Levitski. #31790

sklearn.tree

• Fix Make tree.export_text thread-safe. By Olivier Grisel. #30041

• Fix export_graphviz now raises a ValueError if given feature names are not all strings. By Guilherme Peixoto #31036

• Fix Fixed a regression in decision trees where almost constant features were not handled properly. By Sercan Turkmen. #32259

• Fix Fix handling of missing values in method decision_path of trees (ensemble.DecisionTreeClassifier, ensemble.DecisionTreeRegressor, ensemble.ExtraTreeClassifier and ensemble.ExtraTreeRegressor) By Arthur Lacote. #32280

sklearn.utils

• Efficiency The function sklearn.utils.extmath.safe_sparse_dot was improved by a dedicated Cython routine for the case of a @ b with sparse 2-dimensional a and b and when a dense output is required, i.e., dense_output=True. This improves several algorithms in scikit-learn when dealing with sparse arrays (or matrices). By Christian Lorentzen. #31952

• Enhancement The parameter table in the HTML representation of all scikit-learn estimators and more generally of estimators inheriting from base.BaseEstimator now displays the parameter description as a tooltip and has a link to the online documentation for each parameter. By Dea María Léon. #31564

• Enhancement sklearn.utils._check_sample_weight now raises a clearer error message when the provided weights are neither a scalar nor a 1-D array-like of the same size as the input data. By Kapil Parekh. #31873

• Enhancement sklearn.utils.estimator_checks.parametrize_with_checks now lets you configure strict mode for xfailing checks. Tests that unexpectedly pass will lead to a test failure. The default behaviour is unchanged. By Tim Head. #31951

• Enhancement Fixed the alignment of the “?” and “i” symbols and improved the color style of the HTML representation of estimators. By Guillaume Lemaitre. #31969

• Fix Changes the way color are chosen when displaying an estimator as an HTML representation. Colors are not adapted anymore to the user’s theme, but chosen based on theme declared color scheme (light or dark) for VSCode and JupyterLab. If theme does not declare a color scheme, scheme is chosen according to default text color of the page, if it fails fallbacks to a media query. By Matt J.. #32330

• API Change :function:`utils.extmath.stable_cumsum` is deprecated and will be removed in v1.10. Use np.cumulative_sum with the desired dtype directly instead. By Tiziano Zito #32258. #32258

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.