0.6.0 - 2020-06-09
- Added a new base class called
SupervisedTransformer from which supervised transformers inherit from. Before this, supervised transformers has a
compose.SelectType, which allows selecting feature subsets based on their type.
- Added a
score_one method to
compose.Pipeline so that estimators from the
anomaly module can be pipelined.
compose.Grouper, which allows applying transformers within different subgroups.
datasets.Music, which is a dataset for multi-output binary classification.
datasets.synth.Friedman, which is synthetic regression dataset.
datasets.gen module has been renamed to
- Each dataset now has a
__repr__ method which displays some descriptive information.
datasets.Insects, which has 10 variants.
feature_extraction.Differ has been deprecated. We might put it back in a future if we find a better design.
impute.StatImputer has been completely refactored.
metrics.SMAPE, instead of raising a
ZeroDivisionError, the convention is now to use 0 when both
y_pred are equal to 0.
- Added the possibility to configure how the progress is printed in
model_selection.progressive_val_score. For instance, the progress can now be printed to a file by providing the
- Fixed a bug where
multioutput.RegressorChain could not be pickled.
stats.Shift, which can be used to compute statistics over a shifted version of a variable.
stats.Link, which can be used to compose univariate statistics. Univariate statistics can now be composed via the
stats.RollingCov, which computes covariance between two variables over a window.
stats.RollingPearsonCorr, which computes the Pearson correlation over a window.
- Added a
stream.iter_sql utility method to work with SQLAlchemy.
target_name parameter of
stream.iter_csv has been renamed to
target. It can now be passed a list of values in order to support multi-output scenarios.
stream.iter_arff for handling ARFF files.
- Cancelled the behavior where
tree.DecisionTreeRegressor would raise an exception when no split was found.