0.4.1 - 2019-10-23
base
Tests are now much more extensive, thanks mostly to the newly added estimator tags.
compose
datasets
Added fetch_kdd99_http.
Added fetch_sms.
Added fetch_trec07p.
ensemble
Removed ensemble.HedgeBinaryClassifier because it's performance was subpar.
Removed ensemble.GroupRegressor, as this should be a special case of ensemble.StackingRegressor.
Fixed a bug where feature_extraction.CountVectorizer and feature_extraction.TFIDFVectorizer couldn't be pickled.
linear_model
linear_model.LogisticRegression and linear_model.LinearRegression now have an intercept_lr parameter.
metrics
Metrics can now be composed using the + operator, which is useful for evaluating multiple metrics at the same time.
Added metrics.Rolling, which eliminates the need for a specific rolling implementation for each metric.
Each metric can now be passed a sample_weight argument.
Added metrics.WeightedF1.
Added metrics.WeightedFBeta.
Added metrics.WeightedPrecision.
Added metrics.WeightedRecall.
neighbors
Added neighbors.KNeighborsRegressor.
Added neighbors.KNeighborsClassifier.
optim
Added optim.AdaMax.
The optim module has been reorganized into submodules; namely optim.schedulers, optim.initializers, and optim.losses. The top-level now only contains optimizers. Some classes have been renamed accordingly. See the documentation for details.
Renamed optim.VanillaSGD to optim.SGD.
stats
Added stats.IQR.
Added stats.RollingIQR.
Cythonized stats.Mean and stats.Var.
stream
Added stream.shuffle.
stream.iter_csv now has fraction and seed parameters to sample rows, deterministically or not.Renamed stream.iter_numpy to stream.iter_array.
stream.iter_csv can now read from gzipped files.
time_series
time_series.Detrender now has a window_size parameter for detrending with a rolling mean.
tree
Added tree.RandomForestClassifier.
utils
Fixed a bug where utils.dot could take longer than necessary.