Fix Fixed a bug in cluster.KMeans where KMeans++ initialisation could rarely result in an IndexError. #11756 by Joel Nothman.

Fix Fixed an issue in compose.ColumnTransformer where using DataFrames whose column order differs between :func:fit and :func:transform could lead to silently passing incorrect columns to the remainder transformer. #14237 by Andreas Schuderer <schuderer>.

Fix Fixed a bug in cross_decomposition.CCA improving numerical stability when Y is close to zero. #13903 by Thomas Fan.

Fix Fixed a bug where model_selection.StratifiedKFold shuffles each class’s samples with the same random_state, making shuffle=True ineffective. #13124 by Hanmin Qin.

Fix Fixed a bug in neighbors.KernelDensity which could not be restored from a pickle if sample_weight had been used. #13772 by Aditya Vyas.

Fix Fixed a bug in cluster.KMeans where computation was single threaded when n_jobs > 1 or n_jobs = -1. #12949 by Prabakaran Kumaresshan.

Fix Fixed a bug in compose.ColumnTransformer to handle negative indexes in the columns list of the transformers. #12946 by Pierre Tallotte.

Fix Fixed a regression in covariance.graphical_lasso so that the case n_features=2 is handled correctly. #13276 by Aurélien Bellet.

Fix Fixed a bug in decomposition.sparse_encode where computation was single threaded when n_jobs > 1 or n_jobs = -1. #13005 by Prabakaran Kumaresshan.

Efficiency sklearn.datasets.fetch_openml now loads data by streaming, avoiding high memory usage. #13312 by Joris Van den Bossche.

Fix Fixed a bug in feature_extraction.text.CountVectorizer which would result in the sparse feature matrix having conflicting indptr and indices precisions under very large vocabularies. #11295 by Gabriel Vacaliuc.

Fix add support for non-numeric data in sklearn.impute.MissingIndicator which was not supported while sklearn.impute.SimpleImputer was supporting this for some imputation strategies. #13046 by Guillaume Lemaitre.

Fix Fixed a bug in linear_model.MultiTaskElasticNet and linear_model.MultiTaskLasso which were breaking when warm_start = True. #12360 by Aakanksha Joshi.

Fix Fixed a bug in preprocessing.KBinsDiscretizer where strategy='kmeans' fails with an error during transformation due to unsorted bin edges. #13134 by Sandro Casagrande.

Fix Fixed a bug in preprocessing.OneHotEncoder where the deprecation of categorical_features was handled incorrectly in combination with handle_unknown='ignore'. #12881 by Joris Van den Bossche.

Fix Bins whose width are too small (i.e., <= 1e-8) are removed with a warning in preprocessing.KBinsDiscretizer. #13165 by Hanmin Qin.

Fix Fixed a bug in svm.SVC, svm.NuSVC, svm.SVR, svm.NuSVR and svm.OneClassSVM where the scale option of parameter gamma is erroneously defined as 1 / (n_features * X.std()). It’s now defined as 1 / (n_features * X.var()). #13221 by Hanmin Qin.

Fix Fixed an issue in compose.make_column_transformer which raises unexpected error when columns is pandas Index or pandas Series. #12704 by Hanmin Qin.

Fix Fixed a bug in metrics.pairwise_distances and metrics.pairwise_distances_chunked where parameters V of "seuclidean" and VI of "mahalanobis" metrics were computed after the data was split into chunks instead of being pre-computed on whole data. #12701 by Jeremie du Boisberranger.

Fix Fixed sklearn.neighbors.DistanceMetric jaccard distance function to return 0 when two all-zero vectors are compared. #12685 by Thomas Fan.

Fix Calling utils.check_array on pandas.Series with categorical data, which raised an error in 0.20.0, now returns the expected output again. #12699 by Joris Van den Bossche.

Efficiency make cluster.MeanShift no longer try to do nested parallelism as the overhead would hurt performance significantly when n_jobs > 1. #12159 by Olivier Grisel.

Fix Fixed a bug in cluster.DBSCAN with precomputed sparse neighbors graph, which would add explicitly zeros on the diagonal even when already present. #12105 by Tom Dupre la Tour.

Fix Fixed an issue in compose.ColumnTransformer when stacking columns with types not convertible to a numeric. #11912 by Adrin Jalali.

API Change compose.ColumnTransformer now applies the sparse_threshold even if all transformation results are sparse. #12304 by Andreas Müller.

API Change compose.make_column_transformer now expects (transformer, columns) instead of (columns, transformer) to keep consistent with compose.ColumnTransformer. #12339 by Adrin Jalali.

Fix datasets.fetch_openml to correctly use the local cache. #12246 by Jan N. van Rijn.

Fix datasets.fetch_openml to correctly handle ignore attributes and row id attributes. #12330 by Jan N. van Rijn.

Fix Fixed integer overflow in datasets.make_classification for values of n_informative parameter larger than 64. #10811 by Roman Feldbauer.

Fix Fixed olivetti faces dataset DESCR attribute to point to the right location in datasets.fetch_olivetti_faces. #12441 by Jérémie du Boisberranger

Fix datasets.fetch_openml to retry downloading when reading from local cache fails. #12517 by Thomas Fan.

Fix Fixed a regression in decomposition.IncrementalPCA where 0.20.0 raised an error if the number of samples in the final batch for fitting IncrementalPCA was smaller than n_components. #12234 by Ming Li.

Fix Fixed a bug mostly affecting ensemble.RandomForestClassifier where class_weight='balanced_subsample' failed with more than 32 classes. #12165 by Joel Nothman.

Fix Fixed a bug affecting ensemble.BaggingClassifier, ensemble.BaggingRegressor and ensemble.IsolationForest, where max_features was sometimes rounded down to zero. #12388 by Connor Tann.

Fix Fixed a regression in v0.20.0 where feature_extraction.text.CountVectorizer and other text vectorizers could error during stop words validation with custom preprocessors or tokenizers. #12393 by Roman Yurchak.

Fix linear_model.SGDClassifier and variants with early_stopping=True would not use a consistent validation split in the multiclass case and this would cause a crash when using those estimators as part of parallel parameter search or cross-validation. #12122 by Olivier Grisel.

Fix Fixed a bug affecting SGDClassifier in the multiclass case. Each one-versus-all step is run in a joblib.Parallel call and mutating a common parameter, causing a segmentation fault if called within a backend using processes and not threads. We now use require=sharedmem at the joblib.Parallel instance creation. #12518 by Pierre Glaser and Olivier Grisel.

Fix Fixed a bug in metrics.pairwise.pairwise_distances_argmin_min which returned the square root of the distance when the metric parameter was set to “euclidean”. #12481 by Jérémie du Boisberranger.

Fix Fixed a bug in metrics.pairwise.pairwise_distances_chunked which didn’t ensure the diagonal is zero for euclidean distances. #12612 by Andreas Müller.

API Change The metrics.calinski_harabaz_score has been renamed to metrics.calinski_harabasz_score and will be removed in version 0.23. #12211 by Lisa Thomas, Mark Hannel and Melissa Ferrari.

Fix Ensure that the fit_predict method of mixture.GaussianMixture and mixture.BayesianGaussianMixture always yield assignments consistent with fit followed by predict even if the convergence criterion is too loose or not met. #12451 by Olivier Grisel.

Fix force the parallelism backend to threading for neighbors.KDTree and neighbors.BallTree in Python 2.7 to avoid pickling errors caused by the serialization of their methods. #12171 by Thomas Moreau.

Fix Fixed bug in preprocessing.OrdinalEncoder when passing manually specified categories. #12365 by Joris Van den Bossche.

Fix Fixed bug in preprocessing.KBinsDiscretizer where the transform method mutates the _encoder attribute. The transform method is now thread safe. #12514 by Hanmin Qin.

Fix Fixed a bug in preprocessing.PowerTransformer where the Yeo-Johnson transform was incorrect for lambda parameters outside of [0, 2] #12522 by Nicolas Hug.

Fix Fixed a bug in preprocessing.OneHotEncoder where transform failed when set to ignore unknown numpy strings of different lengths #12471 by Gabriel Marzinotto.

API Change The default value of the method argument in preprocessing.power_transform will be changed from box-cox to yeo-johnson to match preprocessing.PowerTransformer in version 0.23. A FutureWarning is raised when the default value is used. #12317 by Eric Chang.

Fix Use float64 for mean accumulator to avoid floating point precision issues in preprocessing.StandardScaler and decomposition.IncrementalPCA when using float32 datasets. #12338 by bauks.

Fix Calling utils.check_array on pandas.Series, which raised an error in 0.20.0, now returns the expected output again. #12625 by Andreas Müller

Fix When using site joblib by setting the environment variable SKLEARN_SITE_JOBLIB, added compatibility with joblib 0.11 in addition to 0.12+. #12350 by Joel Nothman and Roman Yurchak.

Fix Make sure to avoid raising FutureWarning when calling np.vstack with numpy 1.16 and later (use list comprehensions instead of generator expressions in many locations of the scikit-learn code base). #12467 by Olivier Grisel.

API Change Removed all mentions of sklearn.externals.joblib, and deprecated joblib methods exposed in sklearn.utils, except for utils.parallel_backend and utils.register_parallel_backend, which allow users to configure parallel computation in scikit-learn. Other functionalities are part of joblib. package and should be used directly, by installing it. The goal of this change is to prepare for unvendoring joblib in future version of scikit-learn. #12345 by Thomas Moreau

Major Feature cluster.AgglomerativeClustering now supports Single Linkage clustering via linkage='single'. #9372 by Leland McInnes and Steve Astels.

Feature cluster.KMeans and cluster.MiniBatchKMeans now support sample weights via new parameter sample_weight in fit function. #10933 by Johannes Hansen.

Efficiency cluster.KMeans, cluster.MiniBatchKMeans and cluster.k_means passed with algorithm='full' now enforces row-major ordering, improving runtime. #10471 by Gaurav Dhingra.

Efficiency cluster.DBSCAN now is parallelized according to n_jobs regardless of algorithm. #8003 by Joël Billaud.

Enhancement cluster.KMeans now gives a warning if the number of distinct clusters found is smaller than n_clusters. This may occur when the number of distinct points in the data set is actually smaller than the number of cluster one is looking for. #10059 by Christian Braune.

Fix Fixed a bug where the fit method of cluster.AffinityPropagation stored cluster centers as 3d array instead of 2d array in case of non-convergence. For the same class, fixed undefined and arbitrary behavior in case of training data where all samples had equal similarity. #9612. By Jonatan Samoocha.

Fix Fixed a bug in cluster.spectral_clustering where the normalization of the spectrum was using a division instead of a multiplication. #8129 by Jan Margeta, Guillaume Lemaitre, and Devansh D..

Fix Fixed a bug in cluster.k_means_elkan where the returned iteration was 1 less than the correct value. Also added the missing n_iter_ attribute in the docstring of cluster.KMeans. #11353 by Jeremie du Boisberranger.

Fix Fixed a bug in cluster.mean_shift where the assigned labels were not deterministic if there were multiple clusters with the same intensities. #11901 by Adrin Jalali.

API Change Deprecate pooling_func unused parameter in cluster.AgglomerativeClustering. #9875 by Kumar Ashutosh.

New module.

Major Feature Added compose.ColumnTransformer, which allows to apply different transformers to different columns of arrays or pandas DataFrames. #9012 by Andreas Müller and Joris Van den Bossche, and #11315 by Thomas Fan.

Major Feature Added the compose.TransformedTargetRegressor which transforms the target y before fitting a regression model. The predictions are mapped back to the original space via an inverse transform. #9041 by Andreas Müller and Guillaume Lemaitre.

Efficiency Runtime improvements to covariance.GraphicalLasso. #9858 by Steven Brown.

API Change The covariance.graph_lasso, covariance.GraphLasso and covariance.GraphLassoCV have been renamed to covariance.graphical_lasso, covariance.GraphicalLasso and covariance.GraphicalLassoCV respectively and will be removed in version 0.22. #9993 by Artiem Krinitsyn

Major Feature Added datasets.fetch_openml to fetch datasets from OpenML. OpenML is a free, open data sharing platform and will be used instead of mldata as it provides better service availability. #9908 by Andreas Müller and Jan N. van Rijn.

Feature In datasets.make_blobs, one can now pass a list to the n_samples parameter to indicate the number of samples to generate per cluster. #8617 by Maskani Filali Mohamed and Konstantinos Katrioplas.

Feature Add filename attribute to datasets that have a CSV file. #9101 by alex-33 and Maskani Filali Mohamed.

Feature return_X_y parameter has been added to several dataset loaders. #10774 by Chris Catalfo.

Fix Fixed a bug in datasets.load_boston which had a wrong data point. #10795 by Takeshi Yoshizawa.

Fix Fixed a bug in datasets.load_iris which had two wrong data points. #11082 by Sadhana Srinivasan and Hanmin Qin.

Fix Fixed a bug in datasets.fetch_kddcup99, where data were not properly shuffled. #9731 by Nicolas Goix.

Fix Fixed a bug in datasets.make_circles, where no odd number of data points could be generated. #10045 by Christian Braune.

API Change Deprecated sklearn.datasets.fetch_mldata to be removed in version 0.22. mldata.org is no longer operational. Until removal it will remain possible to load cached datasets. #11466 by Joel Nothman.

Feature decomposition.dict_learning functions and models now support positivity constraints. This applies to the dictionary and sparse code. #6374 by John Kirkham.

Feature Fix decomposition.SparsePCA now exposes normalize_components. When set to True, the train and test data are centered with the train mean repsectively during the fit phase and the transform phase. This fixes the behavior of SparsePCA. When set to False, which is the default, the previous abnormal behaviour still holds. The False value is for backward compatibility and should not be used. #11585 by Ivan Panico.

Efficiency Efficiency improvements in decomposition.dict_learning. #11420 and others by John Kirkham.

Fix Fix for uninformative error in decomposition.IncrementalPCA: now an error is raised if the number of components is larger than the chosen batch size. The n_components=None case was adapted accordingly. #6452. By Wally Gauze.

Fix Fixed a bug where the partial_fit method of decomposition.IncrementalPCA used integer division instead of float division on Python 2. #9492 by James Bourbeau.

Fix In decomposition.PCA selecting a n_components parameter greater than the number of samples now raises an error. Similarly, the n_components=None case now selects the minimum of n_samples and n_features. #8484 by Wally Gauze.

Fix Fixed a bug in decomposition.PCA where users will get unexpected error with large datasets when n_components='mle' on Python 3 versions. #9886 by Hanmin Qin.

Fix Fixed an underflow in calculating KL-divergence for decomposition.NMF #10142 by Tom Dupre la Tour.

Fix Fixed a bug in decomposition.SparseCoder when running OMP sparse coding in parallel using read-only memory mapped datastructures. #5956 by Vighnesh Birodkar and Olivier Grisel.

Efficiency Memory usage improvement for _class_means and _class_cov in discriminant_analysis. #10898 by Nanxin Chen.

Feature dummy.DummyRegressor now has a return_std option in its predict method. The returned standard deviations will be zeros.

Feature dummy.DummyClassifier and dummy.DummyRegressor now only require X to be an object with finite length or shape. #9832 by Vrishank Bhardwaj.

Feature dummy.DummyClassifier and dummy.DummyRegressor can now be scored without supplying test samples. #11951 by Rüdiger Busche.

Feature ensemble.BaggingRegressor and ensemble.BaggingClassifier can now be fit with missing/non-finite values in X and/or multi-output Y to support wrapping pipelines that perform their own imputation. #9707 by Jimmy Wan.

Feature ensemble.GradientBoostingClassifier and ensemble.GradientBoostingRegressor now support early stopping via n_iter_no_change, validation_fraction and tol. #7071 by Raghav RV

Feature Added named_estimators_ parameter in ensemble.VotingClassifier to access fitted estimators. #9157 by Herilalaina Rakotoarison.

Fix Fixed a bug when fitting ensemble.GradientBoostingClassifier or ensemble.GradientBoostingRegressor with warm_start=True which previously raised a segmentation fault due to a non-conversion of CSC matrix into CSR format expected by decision_function. Similarly, Fortran-ordered arrays are converted to C-ordered arrays in the dense case. #9991 by Guillaume Lemaitre.

Fix Fixed a bug in ensemble.GradientBoostingRegressor and ensemble.GradientBoostingClassifier to have feature importances summed and then normalized, rather than normalizing on a per-tree basis. The previous behavior over-weighted the Gini importance of features that appear in later stages. This issue only affected feature importances. #11176 by Gil Forsyth.

API Change The default value of the n_estimators parameter of ensemble.RandomForestClassifier, ensemble.RandomForestRegressor, ensemble.ExtraTreesClassifier, ensemble.ExtraTreesRegressor, and ensemble.RandomTreesEmbedding will change from 10 in version 0.20 to 100 in 0.22. A FutureWarning is raised when the default value is used. #11542 by Anna Ayzenshtat.

API Change Classes derived from ensemble.BaseBagging. The attribute estimators_samples_ will return a list of arrays containing the indices selected for each bootstrap instead of a list of arrays containing the mask of the samples selected for each bootstrap. Indices allows to repeat samples while mask does not allow this functionality. #9524 by Guillaume Lemaitre.

Fix ensemble.BaseBagging where one could not deterministically reproduce fit result using the object attributes when random_state is set. #9723 by Guillaume Lemaitre.

Feature Enable the call to get_feature_names in unfitted feature_extraction.text.CountVectorizer initialized with a vocabulary. #10908 by Mohamed Maskani.

Enhancement idf_ can now be set on a feature_extraction.text.TfidfTransformer. #10899 by Sergey Melderis.

Fix Fixed a bug in feature_extraction.image.extract_patches_2d which would throw an exception if max_patches was greater than or equal to the number of all possible patches rather than simply returning the number of possible patches. #10101 by Varun Agrawal

Fix Fixed a bug in feature_extraction.text.CountVectorizer, feature_extraction.text.TfidfVectorizer, feature_extraction.text.HashingVectorizer to support 64 bit sparse array indexing necessary to process large datasets with more than 2·10⁹ tokens (words or n-grams). #9147 by Claes-Fredrik Mannby and Roman Yurchak.

Fix Fixed bug in feature_extraction.text.TfidfVectorizer which was ignoring the parameter dtype. In addition, feature_extraction.text.TfidfTransformer will preserve dtype for floating and raise a warning if dtype requested is integer. #10441 by Mayur Kulkarni and Guillaume Lemaitre.

Feature Added select K best features functionality to feature_selection.SelectFromModel. #6689 by Nihar Sheth and Quazi Rahman.

Feature Added min_features_to_select parameter to feature_selection.RFECV to bound evaluated features counts. #11293 by Brent Yi.

Feature feature_selection.RFECV’s fit method now supports groups. #9656 by Adam Greenhall.

Fix Fixed computation of n_features_to_compute for edge case with tied CV scores in feature_selection.RFECV. #9222 by Nick Hoh.

Efficiency In gaussian_process.GaussianProcessRegressor, method predict is faster when using return_std=True in particular more when called several times in a row. #9234 by andrewww and Minghui Liu.

New module, adopting preprocessing.Imputer as impute.SimpleImputer with minor changes (see under preprocessing below).

Major Feature Added impute.MissingIndicator which generates a binary indicator for missing values. #8075 by Maniteja Nandana and Guillaume Lemaitre.

Feature The impute.SimpleImputer has a new strategy, 'constant', to complete missing values with a fixed one, given by the fill_value parameter. This strategy supports numeric and non-numeric data, and so does the 'most_frequent' strategy now. #11211 by Jeremie du Boisberranger.

Fix Fixed a bug in isotonic.IsotonicRegression which incorrectly combined weights when fitting a model to data involving points with identical X values. #9484 by Dallas Card

Feature linear_model.SGDClassifier, linear_model.SGDRegressor, linear_model.PassiveAggressiveClassifier, linear_model.PassiveAggressiveRegressor and linear_model.Perceptron now expose early_stopping, validation_fraction and n_iter_no_change parameters, to stop optimization monitoring the score on a validation set. A new learning rate "adaptive" strategy divides the learning rate by 5 each time n_iter_no_change consecutive epochs fail to improve the model. #9043 by Tom Dupre la Tour.

Feature Add sample_weight parameter to the fit method of linear_model.BayesianRidge for weighted linear regression. #10112 by Peter St. John.

Fix Fixed a bug in logistic.logistic_regression_path to ensure that the returned coefficients are correct when multiclass='multinomial'. Previously, some of the coefficients would override each other, leading to incorrect results in linear_model.LogisticRegressionCV. #11724 by Nicolas Hug.

Fix Fixed a bug in linear_model.LogisticRegression where when using the parameter multi_class='multinomial', the predict_proba method was returning incorrect probabilities in the case of binary outcomes. #9939 by Roger Westover.

Fix Fixed a bug in linear_model.LogisticRegressionCV where the score method always computes accuracy, not the metric given by the scoring parameter. #10998 by Thomas Fan.

Fix Fixed a bug in linear_model.LogisticRegressionCV where the ‘ovr’ strategy was always used to compute cross-validation scores in the multiclass setting, even if 'multinomial' was set. #8720 by William de Vazelhes.

Fix Fixed a bug in linear_model.OrthogonalMatchingPursuit that was broken when setting normalize=False. #10071 by Alexandre Gramfort.

Fix Fixed a bug in linear_model.ARDRegression which caused incorrectly updated estimates for the standard deviation and the coefficients. #10153 by Jörg Döpfert.

Fix Fixed a bug in linear_model.ARDRegression and linear_model.BayesianRidge which caused NaN predictions when fitted with a constant target. #10095 by Jörg Döpfert.

Fix Fixed a bug in linear_model.RidgeClassifierCV where the parameter store_cv_values was not implemented though it was documented in cv_values as a way to set up the storage of cross-validation values for different alphas. #10297 by Mabel Villalba-Jiménez.

Fix Fixed a bug in linear_model.ElasticNet which caused the input to be overridden when using parameter copy_X=True and check_input=False. #10581 by Yacine Mazari.

Fix Fixed a bug in sklearn.linear_model.Lasso where the coefficient had wrong shape when fit_intercept=False. #10687 by Martin Hahn.

Fix Fixed a bug in sklearn.linear_model.LogisticRegression where the multi_class='multinomial' with binary output with warm_start=True #10836 by Aishwarya Srinivasan.

Fix Fixed a bug in linear_model.RidgeCV where using integer alphas raised an error. #10397 by Mabel Villalba-Jiménez.

Fix Fixed condition triggering gap computation in linear_model.Lasso and linear_model.ElasticNet when working with sparse matrices. #10992 by Alexandre Gramfort.

Fix Fixed a bug in linear_model.SGDClassifier, linear_model.SGDRegressor, linear_model.PassiveAggressiveClassifier, linear_model.PassiveAggressiveRegressor and linear_model.Perceptron, where the stopping criterion was stopping the algorithm before convergence. A parameter n_iter_no_change was added and set by default to 5. Previous behavior is equivalent to setting the parameter to 1. #9043 by Tom Dupre la Tour.

Fix Fixed a bug where liblinear and libsvm-based estimators would segfault if passed a scipy.sparse matrix with 64-bit indices. They now raise a ValueError. #11327 by Karan Dhingra and Joel Nothman.

API Change The default values of the solver and multi_class parameters of linear_model.LogisticRegression will change respectively from 'liblinear' and 'ovr' in version 0.20 to 'lbfgs' and 'auto' in version 0.22. A FutureWarning is raised when the default values are used. #11905 by Tom Dupre la Tour and Joel Nothman.

API Change Deprecate positive=True option in linear_model.Lars as the underlying implementation is broken. Use linear_model.Lasso instead. #9837 by Alexandre Gramfort.

API Change n_iter_ may vary from previous releases in linear_model.LogisticRegression with solver='lbfgs' and linear_model.HuberRegressor. For Scipy <= 1.0.0, the optimizer could perform more than the requested maximum number of iterations. Now both estimators will report at most max_iter iterations even if more were performed. #10723 by Joel Nothman.

Efficiency Speed improvements for both ‘exact’ and ‘barnes_hut’ methods in manifold.TSNE. #10593 and #10610 by Tom Dupre la Tour.

Feature Support sparse input in manifold.Isomap.fit. #8554 by Leland McInnes.

Feature manifold.t_sne.trustworthiness accepts metrics other than Euclidean. #9775 by William de Vazelhes.

Fix Fixed a bug in manifold.spectral_embedding where the normalization of the spectrum was using a division instead of a multiplication. #8129 by Jan Margeta, Guillaume Lemaitre, and Devansh D..

API Change Feature Deprecate precomputed parameter in function manifold.t_sne.trustworthiness. Instead, the new parameter metric should be used with any compatible metric including ‘precomputed’, in which case the input matrix X should be a matrix of pairwise distances or squared distances. #9775 by William de Vazelhes.

API Change Deprecate precomputed parameter in function manifold.t_sne.trustworthiness. Instead, the new parameter metric should be used with any compatible metric including ‘precomputed’, in which case the input matrix X should be a matrix of pairwise distances or squared distances. #9775 by William de Vazelhes.

Major Feature Added the metrics.davies_bouldin_score metric for evaluation of clustering models without a ground truth. #10827 by Luis Osa.

Major Feature Added the metrics.balanced_accuracy_score metric and a corresponding 'balanced_accuracy' scorer for binary and multiclass classification. #8066 by @xyguo and Aman Dalmia, and #10587 by Joel Nothman.

Feature Partial AUC is available via max_fpr parameter in metrics.roc_auc_score. #3840 by Alexander Niederbühl.

Feature A scorer based on metrics.brier_score_loss is also available. #9521 by Hanmin Qin.

Feature Added control over the normalization in metrics.normalized_mutual_info_score and metrics.adjusted_mutual_info_score via the average_method parameter. In version 0.22, the default normalizer for each will become the arithmetic mean of the entropies of each clustering. #11124 by Arya McCarthy.

Feature Added output_dict parameter in metrics.classification_report to return classification statistics as dictionary. #11160 by Dan Barkhorn.

Feature metrics.classification_report now reports all applicable averages on the given data, including micro, macro and weighted average as well as samples average for multilabel data. #11679 by Alexander Pacha.

Feature metrics.average_precision_score now supports binary y_true other than {0, 1} or {-1, 1} through pos_label parameter. #9980 by Hanmin Qin.

Feature metrics.label_ranking_average_precision_score now supports sample_weight. #10845 by Jose Perez-Parras Toledano.

Feature Add dense_output parameter to metrics.pairwise.linear_kernel. When False and both inputs are sparse, will return a sparse matrix. #10999 by Taylor G Smith.

Efficiency metrics.silhouette_score and metrics.silhouette_samples are more memory efficient and run faster. This avoids some reported freezes and MemoryErrors. #11135 by Joel Nothman.

Fix Fixed a bug in metrics.precision_recall_fscore_support when truncated range(n_labels) is passed as value for labels. #10377 by Gaurav Dhingra.

Fix Fixed a bug due to floating point error in metrics.roc_auc_score with non-integer sample weights. #9786 by Hanmin Qin.

Fix Fixed a bug where metrics.roc_curve sometimes starts on y-axis instead of (0, 0), which is inconsistent with the document and other implementations. Note that this will not influence the result from metrics.roc_auc_score #10093 by alexryndin and Hanmin Qin.

Fix Fixed a bug to avoid integer overflow. Casted product to 64 bits integer in metrics.mutual_info_score. #9772 by Kumar Ashutosh.

Fix Fixed a bug where metrics.average_precision_score will sometimes return nan when sample_weight contains 0. #9980 by Hanmin Qin.

Fix Fixed a bug in metrics.fowlkes_mallows_score to avoid integer overflow. Casted return value of contingency_matrix to int64 and computed product of square roots rather than square root of product. #9515 by Alan Liddell and Manh Dao.

API Change Deprecate reorder parameter in metrics.auc as it’s no longer required for metrics.roc_auc_score. Moreover using reorder=True can hide bugs due to floating point error in the input. #9851 by Hanmin Qin.

API Change In metrics.normalized_mutual_info_score and metrics.adjusted_mutual_info_score, warn that average_method will have a new default value. In version 0.22, the default normalizer for each will become the arithmetic mean of the entropies of each clustering. Currently, metrics.normalized_mutual_info_score uses the default of average_method='geometric', and metrics.adjusted_mutual_info_score uses the default of average_method='max' to match their behaviors in version 0.19. #11124 by Arya McCarthy.

API Change The batch_size parameter to metrics.pairwise_distances_argmin_min and metrics.pairwise_distances_argmin is deprecated to be removed in v0.22. It no longer has any effect, as batch size is determined by global working_memory config. See Limiting Working Memory. #10280 by Joel Nothman and Aman Dalmia.

Feature Added function fit_predict to mixture.GaussianMixture and mixture.GaussianMixture, which is essentially equivalent to calling fit and predict. #10336 by Shu Haoran and Andrew Peng.

Fix Fixed a bug in mixture.BaseMixture where the reported n_iter_ was missing an iteration. It affected mixture.GaussianMixture and mixture.BayesianGaussianMixture. #10740 by Erich Schubert and Guillaume Lemaitre.

Fix Fixed a bug in mixture.BaseMixture and its subclasses mixture.GaussianMixture and mixture.BayesianGaussianMixture where the lower_bound_ was not the max lower bound across all initializations (when n_init > 1), but just the lower bound of the last initialization. #10869 by Aurélien Géron.

Feature Add return_estimator parameter in model_selection.cross_validate to return estimators fitted on each split. #9686 by Aurélien Bellet.

Feature New refit_time_ attribute will be stored in model_selection.GridSearchCV and model_selection.RandomizedSearchCV if refit is set to True. This will allow measuring the complete time it takes to perform hyperparameter optimization and refitting the best model on the whole dataset. #11310 by Matthias Feurer.

Feature Expose error_score parameter in model_selection.cross_validate, model_selection.cross_val_score, model_selection.learning_curve and model_selection.validation_curve to control the behavior triggered when an error occurs in model_selection._fit_and_score. #11576 by Samuel O. Ronsin.

Feature BaseSearchCV now has an experimental, private interface to support customized parameter search strategies, through its _run_search method. See the implementations in model_selection.GridSearchCV and model_selection.RandomizedSearchCV and please provide feedback if you use this. Note that we do not assure the stability of this API beyond version 0.20. #9599 by Joel Nothman

Enhancement Add improved error message in model_selection.cross_val_score when multiple metrics are passed in scoring keyword. #11006 by Ming Li.

API Change The default number of cross-validation folds cv and the default number of splits n_splits in the model_selection.KFold-like splitters will change from 3 to 5 in 0.22 as 3-fold has a lot of variance. #11557 by Alexandre Boucaud.

API Change The default of iid parameter of model_selection.GridSearchCV and model_selection.RandomizedSearchCV will change from True to False in version 0.22 to correspond to the standard definition of cross-validation, and the parameter will be removed in version 0.24 altogether. This parameter is of greatest practical significance where the sizes of different test sets in cross-validation were very unequal, i.e. in group-based CV strategies. #9085 by Laurent Direr and Andreas Müller.

API Change The default value of the error_score parameter in model_selection.GridSearchCV and model_selection.RandomizedSearchCV will change to np.NaN in version 0.22. #10677 by Kirill Zhdanovich.

API Change Changed ValueError exception raised in model_selection.ParameterSampler to a UserWarning for case where the class is instantiated with a greater value of n_iter than the total space of parameters in the parameter grid. n_iter now acts as an upper bound on iterations. #10982 by Juliet Lawton

API Change Invalid input for model_selection.ParameterGrid now raises TypeError. #10928 by Solutus Immensus

Major Feature Added multioutput.RegressorChain for multi-target regression. #9257 by Kumar Ashutosh.

Major Feature Added naive_bayes.ComplementNB, which implements the Complement Naive Bayes classifier described in Rennie et al. (2003). #8190 by Michael A. Alcorn.

Feature Add var_smoothing parameter in naive_bayes.GaussianNB to give a precise control over variances calculation. #9681 by Dmitry Mottl.

Fix Fixed a bug in naive_bayes.GaussianNB which incorrectly raised error for prior list which summed to 1. #10005 by Gaurav Dhingra.

Fix Fixed a bug in naive_bayes.MultinomialNB which did not accept vector valued pseudocounts (alpha). #10346 by Tobias Madsen

Efficiency neighbors.RadiusNeighborsRegressor and neighbors.RadiusNeighborsClassifier are now parallelized according to n_jobs regardless of algorithm. #10887 by Joël Billaud.

Efficiency Nearest neighbors query methods are now more memory efficient when algorithm='brute'. #11136 by Joel Nothman and Aman Dalmia.

Feature Add sample_weight parameter to the fit method of neighbors.KernelDensity to enable weighting in kernel density estimation. #4394 by Samuel O. Ronsin.

Feature Novelty detection with neighbors.LocalOutlierFactor: Add a novelty parameter to neighbors.LocalOutlierFactor. When novelty is set to True, neighbors.LocalOutlierFactor can then be used for novelty detection, i.e. predict on new unseen data. Available prediction methods are predict, decision_function and score_samples. By default, novelty is set to False, and only the fit_predict method is avaiable. By Albert Thomas.

Fix Fixed a bug in neighbors.NearestNeighbors where fitting a NearestNeighbors model fails when a) the distance metric used is a callable and b) the input to the NearestNeighbors model is sparse. #9579 by Thomas Kober.

Fix Fixed a bug so predict in neighbors.RadiusNeighborsRegressor can handle empty neighbor set when using non uniform weights. Also raises a new warning when no neighbors are found for samples. #9655 by Andreas Bjerre-Nielsen.

Fix Efficiency Fixed a bug in KDTree construction that results in faster construction and querying times. #11556 by Jake VanderPlas

Fix Fixed a bug in neighbors.KDTree and neighbors.BallTree where pickled tree objects would change their type to the super class BinaryTree. #11774 by Nicolas Hug.

Feature Add n_iter_no_change parameter in neural_network.BaseMultilayerPerceptron, neural_network.MLPRegressor, and neural_network.MLPClassifier to give control over maximum number of epochs to not meet tol improvement. #9456 by Nicholas Nadeau.

Fix Fixed a bug in neural_network.BaseMultilayerPerceptron, neural_network.MLPRegressor, and neural_network.MLPClassifier with new n_iter_no_change parameter now at 10 from previously hardcoded 2. #9456 by Nicholas Nadeau.

Fix Fixed a bug in neural_network.MLPRegressor where fitting quit unexpectedly early due to local minima or fluctuations. #9456 by Nicholas Nadeau

Feature The predict method of pipeline.Pipeline now passes keyword arguments on to the pipeline’s last estimator, enabling the use of parameters such as return_std in a pipeline with caution. #9304 by Breno Freitas.

API Change pipeline.FeatureUnion now supports 'drop' as a transformer to drop features. #11144 by Thomas Fan.

Major Feature Expanded preprocessing.OneHotEncoder to allow to encode categorical string features as a numeric array using a one-hot (or dummy) encoding scheme, and added preprocessing.OrdinalEncoder to convert to ordinal integers. Those two classes now handle encoding of all feature types (also handles string-valued features) and derives the categories based on the unique values in the features instead of the maximum value in the features. #9151 and #10521 by Vighnesh Birodkar and Joris Van den Bossche.

Major Feature Added preprocessing.KBinsDiscretizer for turning continuous features into categorical or one-hot encoded features. #7668, #9647, #10195, #10192, #11272, #11467 and #11505. by Henry Lin, Hanmin Qin, Tom Dupre la Tour and Giovanni Giuseppe Costa.

Major Feature Added preprocessing.PowerTransformer, which implements the Yeo-Johnson and Box-Cox power transformations. Power transformations try to find a set of feature-wise parametric transformations to approximately map data to a Gaussian distribution centered at zero and with unit variance. This is useful as a variance-stabilizing transformation in situations where normality and homoscedasticity are desirable. #10210 by Eric Chang and Maniteja Nandana, and #11520 by Nicolas Hug.

Major Feature NaN values are ignored and handled in the following preprocessing methods: preprocessing.MaxAbsScaler, preprocessing.MinMaxScaler, preprocessing.RobustScaler, preprocessing.StandardScaler, preprocessing.PowerTransformer, preprocessing.QuantileTransformer classes and preprocessing.maxabs_scale, preprocessing.minmax_scale, preprocessing.robust_scale, preprocessing.scale, preprocessing.power_transform, preprocessing.quantile_transform functions respectively addressed in issues #11011, #11005, #11308, #11206, #11306, and #10437. By Lucija Gregov and Guillaume Lemaitre.

Feature preprocessing.PolynomialFeatures now supports sparse input. #10452 by Aman Dalmia and Joel Nothman.

Feature preprocessing.RobustScaler and preprocessing.robust_scale can be fitted using sparse matrices. #11308 by Guillaume Lemaitre.

Feature preprocessing.OneHotEncoder now supports the get_feature_names method to obtain the transformed feature names. #10181 by Nirvan Anjirbag and Joris Van den Bossche.

Feature A parameter check_inverse was added to preprocessing.FunctionTransformer to ensure that func and inverse_func are the inverse of each other. #9399 by Guillaume Lemaitre.

Feature The transform method of sklearn.preprocessing.MultiLabelBinarizer now ignores any unknown classes. A warning is raised stating the unknown classes classes found which are ignored. #10913 by Rodrigo Agundez.

Fix Fixed bugs in preprocessing.LabelEncoder which would sometimes throw errors when transform or inverse_transform was called with empty arrays. #10458 by Mayur Kulkarni.

Fix Fix ValueError in preprocessing.LabelEncoder when using inverse_transform on unseen labels. #9816 by Charlie Newey.

Fix Fix bug in preprocessing.OneHotEncoder which discarded the dtype when returning a sparse matrix output. #11042 by Daniel Morales.

Fix Fix fit and partial_fit in preprocessing.StandardScaler in the rare case when with_mean=False and with_std=False which was crashing by calling fit more than once and giving inconsistent results for mean_ whether the input was a sparse or a dense matrix. mean_ will be set to None with both sparse and dense inputs. n_samples_seen_ will be also reported for both input types. #11235 by Guillaume Lemaitre.

API Change Deprecate n_values and categorical_features parameters and active_features_, feature_indices_ and n_values_ attributes of preprocessing.OneHotEncoder. The n_values parameter can be replaced with the new categories parameter, and the attributes with the new categories_ attribute. Selecting the categorical features with the categorical_features parameter is now better supported using the compose.ColumnTransformer. #10521 by Joris Van den Bossche.

API Change Deprecate preprocessing.Imputer and move the corresponding module to impute.SimpleImputer. #9726 by Kumar Ashutosh.

API Change The axis parameter that was in preprocessing.Imputer is no longer present in impute.SimpleImputer. The behavior is equivalent to axis=0 (impute along columns). Row-wise imputation can be performed with FunctionTransformer (e.g., FunctionTransformer(lambda X: SimpleImputer().fit_transform(X.T).T)). #10829 by Guillaume Lemaitre and Gilberto Olimpio.

API Change The NaN marker for the missing values has been changed between the preprocessing.Imputer and the impute.SimpleImputer. missing_values='NaN' should now be missing_values=np.nan. #11211 by Jeremie du Boisberranger.

API Change In preprocessing.FunctionTransformer, the default of validate will be from True to False in 0.22. #10655 by Guillaume Lemaitre.

Fix Fixed a bug in svm.SVC where when the argument kernel is unicode in Python2, the predict_proba method was raising an unexpected TypeError given dense inputs. #10412 by Jiongyan Zhang.

API Change Deprecate random_state parameter in svm.OneClassSVM as the underlying implementation is not random. #9497 by Albert Thomas.

API Change The default value of gamma parameter of svm.SVC, NuSVC, SVR, NuSVR, OneClassSVM will change from 'auto' to 'scale' in version 0.22 to account better for unscaled features. #8361 by Gaurav Dhingra and Ting Neo.

Enhancement Although private (and hence not assured API stability), tree._criterion.ClassificationCriterion and tree._criterion.RegressionCriterion may now be cimported and extended. #10325 by Camil Staps.

Fix Fixed a bug in tree.BaseDecisionTree with splitter="best" where split threshold could become infinite when values in X were near infinite. #10536 by Jonathan Ohayon.

Fix Fixed a bug in tree.MAE to ensure sample weights are being used during the calculation of tree MAE impurity. Previous behaviour could cause suboptimal splits to be chosen since the impurity calculation considered all samples to be of equal weight importance. #11464 by John Stott.

Feature utils.check_array and utils.check_X_y now have accept_large_sparse to control whether scipy.sparse matrices with 64-bit indices should be rejected. #11327 by Karan Dhingra and Joel Nothman.

Efficiency Fix Avoid copying the data in utils.check_array when the input data is a memmap (and copy=False). #10663 by Arthur Mensch and Loïc Estève.

API Change utils.check_array yield a FutureWarning indicating that arrays of bytes/strings will be interpreted as decimal numbers beginning in version 0.22. #10229 by Ryan Lee

Feature API Change More consistent outlier detection API: Add a score_samples method in svm.OneClassSVM, ensemble.IsolationForest, neighbors.LocalOutlierFactor, covariance.EllipticEnvelope. It allows to access raw score functions from original papers. A new offset_ parameter allows to link score_samples and decision_function methods. The contamination parameter of ensemble.IsolationForest and neighbors.LocalOutlierFactor decision_function methods is used to define this offset_ such that outliers (resp. inliers) have negative (resp. positive) decision_function values. By default, contamination is kept unchanged to 0.1 for a deprecation period. In 0.22, it will be set to “auto”, thus using method-specific score offsets. In covariance.EllipticEnvelope decision_function method, the raw_values parameter is deprecated as the shifted Mahalanobis distance will be always returned in 0.22. #9015 by Nicolas Goix.

Feature API Change A behaviour parameter has been introduced in ensemble.IsolationForest to ensure backward compatibility. In the old behaviour, the decision_function is independent of the contamination parameter. A threshold attribute depending on the contamination parameter is thus used. In the new behaviour the decision_function is dependent on the contamination parameter, in such a way that 0 becomes its natural threshold to detect outliers. Setting behaviour to “old” is deprecated and will not be possible in version 0.22. Beside, the behaviour parameter will be removed in 0.24. #11553 by Nicolas Goix.

API Change Added convergence warning to svm.LinearSVC and linear_model.LogisticRegression when verbose is set to 0. #10881 by Alexandre Sevin.

API Change Changed warning type from UserWarning to exceptions.ConvergenceWarning for failing convergence in linear_model.logistic_regression_path, linear_model.RANSACRegressor, linear_model.ridge_regression, gaussian_process.GaussianProcessRegressor, gaussian_process.GaussianProcessClassifier, decomposition.fastica, cross_decomposition.PLSCanonical, cluster.AffinityPropagation, and cluster.Birch. #10306 by Jonathan Siebert.

Major Feature A new configuration parameter, working_memory was added to control memory consumption limits in chunked operations, such as the new metrics.pairwise_distances_chunked. See Limiting Working Memory. #10280 by Joel Nothman and Aman Dalmia.

Feature The version of joblib bundled with Scikit-learn is now 0.12. This uses a new default multiprocessing implementation, named loky. While this may incur some memory and communication overhead, it should provide greater cross-platform stability than relying on Python standard library multiprocessing. #11741 by the Joblib developers, especially Thomas Moreau and Olivier Grisel.

Feature An environment variable to use the site joblib instead of the vendored one was added (Environment variables). The main API of joblib is now exposed in sklearn.utils. #11166 by Gael Varoquaux.

Feature Add almost complete PyPy 3 support. Known unsupported functionalities are datasets.load_svmlight_file, feature_extraction.FeatureHasher and feature_extraction.text.HashingVectorizer. For running on PyPy, PyPy3-v5.10+, Numpy 1.14.0+, and scipy 1.1.0+ are required. #11010 by Ronan Lamy and Roman Yurchak.

Feature A utility method sklearn.show_versions was added to print out information relevant for debugging. It includes the user system, the Python executable, the version of the main libraries and BLAS binding information. #11596 by Alexandre Boucaud

Fix Fixed a bug when setting parameters on meta-estimator, involving both a wrapped estimator and its parameter. #9999 by Marcus Voss and Joel Nothman.

Fix Fixed a bug where calling sklearn.base.clone was not thread safe and could result in a “pop from empty list” error. #9569 by Andreas Müller.

API Change The default value of n_jobs is changed from 1 to None in all related functions and classes. n_jobs=None means unset. It will generally be interpreted as n_jobs=1, unless the current joblib.Parallel backend context specifies otherwise (See Glossary for additional information). Note that this change happens immediately (i.e., without a deprecation cycle). #11741 by Olivier Grisel.

Fix Fixed a bug in validation helpers where passing a Dask DataFrame results in an error. #12462 by Zachariah Miller