Version 1.10

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 1.10.dev0

September 2026

Changed models

• preprocessing.QuantileTransformer and preprocessing.quantile_transform now estimate quantiles with the averaged_inverted_cdf method and no longer cap n_quantiles at n_samples. Dense subsampling is also done with replacement. These changes affect quantiles_ even when sample_weight is not passed. By Karassay Raushanbek and Shruti Nath. #32761

• inspection.partial_dependence and inspection.PartialDependenceDisplay now use numpy.nanquantile’s linear interpolation instead of the (unmaintained) scipy.stats.mstats.mquantiles’s Cunnane interpolation to compute the grid of values and deciles from X. This may slightly shift the computed grid values and displayed deciles. np.nan values in a numerical column are now ignored (instead of being sorted last in the quantile calculation and hence possibly affecting the result depending on the requested percentiles). By Arthur Lacote. #34928

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 covariance.LedoitWolf now supports array API compatible inputs. By Tim Head #33573

• Feature covariance.OAS now supports array API compatible inputs. By Bruno Aristimunha #33600

• Feature linear_model.LogisticRegressionCV now supports array API compatible inputs with solver="lbfgs" and solver="newton-cg" and when the underlying scoring function is also compatible with the array API. By Omar Salman and Christian Lorentzen. #33906

• Feature sklearn.metrics.matthews_corrcoef now supports array API compatible inputs. By Mohamad Fazeli. #34423

• Feature sklearn.metrics.cluster.contingency_matrix now supports array API compatible inputs. By Yaroslav Korobko and Omar Salman. #34785

• Enhancement Add support for array inputs from mixed namespaces and devices for: sklearn.metrics.det_curve, sklearn.metrics.roc_curve, sklearn.metrics.zero_one_loss, sklearn.metrics.jaccard_score, sklearn.metrics.balanced_accuracy_score, sklearn.metrics.cohen_kappa_score. By Lucy Liu. #34442

  Enhancement linear_model.LogisticRegression now supports array API compatible inputs with solver="newton-cg". By Christian Lorentzen and Omar Salman. #34412

• Fix model_selection.GridSearchCV, model_selection.RandomizedSearchCV, model_selection.cross_validate and model_selection.cross_val_score no longer force y into X’s array namespace during cross-validation. Namespace and device handling of y is delegated to the wrapped estimator, so an array API X can be combined with a NumPy y, including a NumPy string y which cannot be held by array API namespaces. By Tim Head. #33633

• Fix Fixed metrics.median_absolute_error to handle integer inputs when the input array namespace does not support integer inputs to its quantile function (e.g., PyTorch). By Lucy Liu. #33778

• Fix Fixes support when average="samples" and sample_weight is an array from a namespace or device different from y_pred, for: sklearn.metrics.precision_recall_fscore_support, sklearn.metrics.recall, sklearn.metrics.precision, sklearn.metrics.f1_score, sklearn.metrics.fbeta_score. Also fixes several bugs in array API compliance of sklearn.metrics.roc_curve. By Lucy Liu. #34442

• Fix discriminant_analysis.LinearDiscriminantAnalysis now supports array API inputs where y is from a different namespace or device than X, including string labels. By Kiyarash Fazeli. #34475

• Fix linear_model.PoissonRegressor now accepts y and sample_weight from a different array API namespace or device than X. By Amineh Dadsetan. #34480

• Fix Fixed classification metrics to accept pandas labels when array API dispatch is enabled and predictions are from a different array API namespace or device. By Olivier Grisel. #34779

Metadata routing

Refer to the Metadata Routing User Guide for more details.

• Feature Feature Added support for auto-requesting metadata through set_config(enable_metadata_auto_requests=True). This enables automatic routing of common metadata like sample_weight, or X_val. Consumers (estimators, scorers, and CV splitters) and consuming routers can now define which metadata requests to set automatically via the add_auto_request method, which complements the existing class-level defaults set through __metadata_request__* attributes. Refer to Default and Auto-Requested Metadata for more details. By Adrin Jalali and Stefanie Senger. #31413

Callbacks

• API Change In callback.ScoringMonitor, the option to provide scoring=None to have the default scorer of the estimator has been removed. By François Paugam. #34590

sklearn.calibration

• Feature calibration.calibration_curve, calibration.CalibrationDisplay.from_estimator, and calibration.CalibrationDisplay.from_predictions now support n_bins="cube_root". This option sets the number of bins to ceil(n_samples**(1/3)). By Zeyu Sun #33856

• Fix calibration.CalibratedClassifierCV with method="sigmoid" now converts predict_proba outputs to logits (log(p/(1-p))) before calibration instead of passing probabilities directly to the calibrator. For method="temperature", predict_proba outputs are converted to multinomial logits. Estimators that only expose decision_function are unaffected. Calibration now prefers predict_proba over decision_function when both are available. By Olivier Grisel and Antoine Baker. #34313

sklearn.cluster

• Efficiency cluster.AgglomerativeClustering and cluster.FeatureAgglomeration with linkage="single", as well as cluster.HDBSCAN, are now faster for inputs that produce deep union-find trees. Restored path compression avoids quadratic time while constructing the single-linkage hierarchy. By Chris Boseak. #34872

sklearn.datasets

• Fix datasets.fetch_openml with parser="pandas" no longer raises a TypeError when a nominal column’s values are inferred by pandas as numeric or boolean categories instead of strings. By Dea María Léon. #34902

sklearn.discriminant_analysis

• Efficiency discriminant_analysis.LinearDiscriminantAnalysis is faster to fit: the target is ordinal-encoded once in fit and the number of classes is passed down to the internal helpers, removing a redundant pass over the target to recompute the unique class labels. Computing the class means is up to 1.4x faster for large numbers of samples and classes. By Kiyarash Fazeli. #34475

sklearn.ensemble

• Efficiency Part of the fit of sklearn.ensemble.GradientBoostingClassifier was optimized. The speed-up is mainly visible for deeper trees, which is not the primary use case for gradient boosting but can occur in grid searches. For very deep trees, this can be up to 20 times faster. By Arthur Lacote #32911

• Efficiency Fitting in sklearn.ensemble.HistGradientBoostingClassifier and sklearn.ensemble.HistGradientBoostingRegressor was optimized to be faster, by speeding up assignment of bins. By Itamar Turner-Trauring. #34194

• Efficiency Improved the performance of binning in ensemble.HistGradientBoostingClassifier and ensemble.HistGradientBoostingRegressor. Fitting with sample weights could be slow for datasets with fewer than 200,000 samples; this is now fixed. Binning without sample weights was also optimized. By Arthur Lacote. #34248

• Efficiency Prediction in sklearn.ensemble.HistGradientBoostingClassifier and sklearn.ensemble.HistGradientBoostingRegressor is faster when using categoricals. By Itamar Turner-Trauring. #34486

• Efficiency sklearn.ensemble.RandomForestClassifer training is now faster on sparse datasets. By Itamar Turner-Trauring. #34586

• API Change The sklearn.experimental.enable_hist_gradient_boosting module is deprecated and will be removed in 1.12. It is no longer needed since ensemble.HistGradientBoostingClassifier and ensemble.HistGradientBoostingRegressor are stable and can be imported normally from sklearn.ensemble. By Guillaume Lemaitre. #34236

sklearn.feature_selection

• Fix Feature selectors such as feature_selection.SelectKBest now correctly reject non-finite values (infinity, and NaN unless allowed) in transform when using set_output(transform="pandas"). By Abhimanyu Singh Shekhawat. #34513

sklearn.impute

• API Change impute.IterativeImputer is no longer experimental. It can now be imported directly with from sklearn.impute import IterativeImputer, and importing sklearn.experimental.enable_iterative_imputer is no longer required (doing so now raises a warning and is a no-op). It also no longer raises a exceptions.ConvergenceWarning when max_iter is reached without meeting the tol stopping criterion, as non-convergence of the round-robin imputation is expected. The User Guide now documents why the imputed values are not guaranteed to converge and why investing in better imputation often yields diminishing returns for prediction. By Guillaume Lemaitre. #34214

sklearn.inspection

• Enhancement The parameter multiclass_colors was deprecated in favour of target_colors in inspection.DecisionBoundaryDisplay. The attribute multiclass_colors_ was also renamed to target_colors_. Now they can be used for binary problems as well without causing confusion (which will be added in a follow-up PR). By Anne Beyer. #34092

sklearn.linear_model

• Major Feature The generalized linear models (GLM) linear_model.GammaRegressor, linear_model.PoissonRegressor and linear_model.TweedieRegressor are now able to fit L1 and Elastic-Net penalties with the new parameter l1_ratio together with the 2 new solvers "newton-cd" and "newton-cd-gram". These new solvers are also available for linear_model.LogisticRegression. They are based on the Newton solver infrastructure introduced for "newton-cholesky" and take advantage of the already available and fast (e.g. via gap safe screening rules) coordinate descent solvers of ElasticNet.

  • "newton-cd-gram" constructs the full Hessian (gram) matrix and then uses ElasticNet(precompute=True) as the inner solver. It is a good choice for n_samples >> n_features.

  • "newton-cd" avoids constructing the full Hessian matrix, but formulates the minimization problem for the inner solver as a penalized least squares problem (which amounts to taking the square root of the Hessian) and then uses ElasticNet(precompute=False) on it. It is a good choice for n_features > n_samples.

    By Christian Lorentzen. #34377, #34523. #34561

• Efficiency The "newton-cholesky" solver of linear_model.LogisticRegression, linear_model.GammaRegressor, linear_model.PoissonRegressor and linear_model.TweedieRegressor now uses an improved backtracking line search. Instead of halving the step size in each line search iteration, it uses cubic or quadratic polynomial interpolation to find the next trial step size. By Christian Lorentzen. #34157

• Efficiency All solvers of linear_model.LogisticRegressionCV are now faster on free-threaded Python when using parallelism. The only exceptions are "sag" and "saga" solvers, which will continue to run at the same speed. By Itamar Turner-Trauring. #34659

• Enhancement linear_model.PoissonRegressor, linear_model.GammaRegressor and linear_model.TweedieRegressor got a new solver: solver="newton-cg". This is the same Newton conjugate gradient solver, a.k.a. truncated Newton, that is already available with linear_model.LogisticRegression. By Christian Lorentzen. #33759

• Enhancement linear_model.ElasticNetCV, linear_model.LassoCV and linear_model.MultiTaskElasticNetCV as well as enet_path and lasso_path are now faster. The (relative) speed-up is larger for lower tolerance (larger values of tol) and achieved by avoiding to re-calculate the residuals and the column norms of X for each penalty alpha along the path. By Christian Lorentzen. #34572

• Enhancement linear_model.ElasticNetCV, linear_model.LassoCV and linear_model.MultiTaskElasticNetCV now avoid prematurely converting X to F-contiguous array or sparse CSC because the inner cross validation take subsamples of X which then need the same conversion again, anyway. So this PR can avoid memory copies of X. By Christian Lorentzen. #34613

• Fix The solver="newton-cg" for linear_model.LogisticRegression, linear_model.PoissonRegressor and other GLMs is now more robust. Many convergence issues have been fixed, e.g. for certain ill-conditioned problems with low regularization and nearly collinear features. The fix is about treating small or even negative curvature. By Christian Lorentzen. #34412

sklearn.metrics

• Efficiency metrics.nan_euclidean_distances is now several times faster on dense data, with a more moderate speed-up for impute.KNNImputer and estimators using metric="nan_euclidean" such as neighbors.NearestNeighbors. By Roman Yurchak. #1140

• Enhancement confusion_matrix_at_thresholds, roc_curve, precision_recall_curve, det_curve, and roc_auc_score now use exact integer cumulative sums for unweighted inputs (sample_weight=None), preventing floating-point precision saturation on float32-only array API devices when processing datasets exceeding \(2^{24}\) samples. By Kaynup. #34817

• Fix Fixed metrics.average_precision_score to handle list-type y_score for multiclass data. By Lucy Liu. #34083

• API Change The default value of the average parameter of precision_recall_fscore_support will change from None to "binary" in version 1.12. By François Paugam. #34190

• API Change In metrics.DetCurveDisplay the estimator_name parameter is deprecated in favour of name and will be removed in 1.12. The **kwargs parameter of metrics.DetCurveDisplay.plot, metrics.DetCurveDisplay.from_estimator and metrics.DetCurveDisplay.from_predictions is also deprecated in favour of curve_kwargs and will be removed in 1.12. By @AnneBeyer. #34443

sklearn.mixture

• Efficiency mixture.GaussianMixture with covariance_type="tied" is now faster, with the speed-up most visible for many components or high-dimensional data. By Roman Yurchak. #1140

sklearn.model_selection

• API Change model_selection.HalvingGridSearchCV and model_selection.HalvingRandomSearchCV now expose the cross-validation results of the last halving iteration in cv_results_. The full per-iteration history is available in the new all_cv_results_ attribute. By Guillaume Lemaitre. #34250

sklearn.neighbors

• Feature neighbors.NeighborhoodComponentsAnalysis now supports sparse input matrices for init in ['pca', 'random', 'identity']. By Arturo Amor. #34122

• Efficiency neighbors.KNeighborsRegressor.predict is now faster for regression with many output targets. By Roman Yurchak. #1140

• Efficiency sklearn.neighbors.KNeighborsClassifier is now faster when running with multiple threads. The statistics returned from sklearn.neighbors.BallTree and sklearn.neighbors.KDTree no longer reset on each query, and are no longer incorrect when the tree is used from multiple threads. By Itamar Turner-Trauring. #34187

• Efficiency sklearn.neighbors.KNeighborsClassifier is now faster when fitting when used with sklearn.neighbors.BallTree or sklearn.neighbors.KDTree. By Itamar Turner-Trauring. #34564

• API Change Getting statistics and number of calls from sklearn.neighbors.BallTree and sklearn.neighbors.KDTree is deprecated. The relevant APIs will be removed in version 1.12. #34242

sklearn.pipeline

• Fix The default value for the transform_input parameter of Pipeline was changed from None to ("X_val",) so that the validation set, when passed to fit, is always transformed alongside X, to prevent easy to miss mistakes. By Jérémie du Boisberranger. #34263

sklearn.preprocessing

• Efficiency Improved fit time of preprocessing.OneHotEncoder and preprocessing.OrdinalEncoder on object/string categorical features when category counts are needed, for instance with min_frequency or max_categories. By Itamar Turner-Trauring. #34386

• Efficiency preprocessing.OneHotEncoder, preprocessing.OrdinalEncoder, and preprocessing.TargetEncoder are now faster by using a Fortran-contiguous memory layout in their encoding paths. By Arthur Lacote. #34392

• Enhancement preprocessing.QuantileTransformer can now handle sample weight values by subsampling using sample weights if subsampling is enabled or otherwise using _weighted_percentile. The default method of calculating percentiles for both weighted and unweighted cases is now fixed to “averaged_inverted_cdf”. Support is only implemented for dense X. By Karassay Raushanbek and Shruti Nath. #32761

• Fix A few edge cases with missing values and unique knot values in preprocessing.SplineTransformer are now handled correctly. For instance, if a feature is constant, i.e., has only one unique value at fit time, transform now always returns all zeros for that feature. By Christian Lorentzen. #33953

• Fix The warning raised by preprocessing.OneHotEncoder now distinguishes columns where unknown categories are encoded as the infrequent category from columns where they are encoded as all zeros. By oathbreaker. #34861

sklearn.tree

• Major Feature tree.DecisionTreeRegressor, tree.DecisionTreeClassifier now have native support for categorical features for binary classification and single-output regression. All criteria are supported except for 'absolute_error'. Categorical features can be specified with the categorical_features parameter. Up to 256 categories per features are supported. By Adam Li, Arthur Lacote, Adrin Jalali and Christian Lorentzen #33354

• Major Feature tree.DecisionTreeRegressor, tree.DecisionTreeClassifier, tree.ExtraTreeRegressor, and tree.ExtraTreeClassifier now have native support for categorical features via the categorical_features parameter. Decision trees support binary classification and single-output regression (all criteria except 'absolute_error') with up to 256 categories per feature. Extra trees also support multi-class / multi-output targets with up to 2**24 - 1 categories via random hash splits. By Adam Li, Arthur Lacote, Adrin Jalali and Christian Lorentzen #33972

• Feature Added a fill_colors parameter to tree.plot_tree and tree.export_graphviz to allow user-defined class colors when rendering filled decision trees. By Simon-Martin Schröder. #33810

Code and documentation contributors

Thanks to everyone who has contributed to the maintenance and improvement of the project since version 1.9, including:

TODO: update at the time of the release.