0.5.0 - 2020-03-13

• PyPI
• GitHub

compat

• Added compat.PyTorch2CremeRegressor.
• compat.SKL2CremeRegressor and compat.SKL2CremeClassifier now have an optional batch_size parameter in order to perform mini-batching.

compose

• Renamed compose.Whitelister to compose.Select.
• Renamed compose.Blacklister to compose.Discard.

facto

• Added facto.FFMClassifier.
• Added facto.FFMRegressor.
• Added facto.FwFMClassifier.
• Added facto.FwFMRegressor.
• Added facto.HOFMClassifier.
• Added facto.HOFMRegressor.
• Refactored facto.FMClassifier.
• Refactored facto.FMRegressor.

feature_selection

• Added feature_selection.PoissonInclusion.
• Removed feature_selection.RandomDiscarder as it didn't make much sense.

feature_extraction

• Renamed feature_extraction.CountVectorizer to feature_extraction.BagOfWords.
• Renamed feature_extraction.TFIDFVectorizer to feature_extraction.TFIDF.
• Added preprocessor and ngram_range parameters to feature_extraction.BagOfWords.
• Added preprocessor and ngram_range parameters to feature_extraction.TFIDF.

datasets

• The datasets module has been overhauled. Each dataset is now a class (e.g. fetch_electricity has become datasets.Elec2).
• Added datasets.TrumpApproval.
• Added datasets.MaliciousURL.
• Added datasets.gen.SEA.
• Added datasets.Higgs.
• Added datasets.MovieLens100K.
• Added datasets.Bananas.
• Added datasets.Taxis.
• Added datasets.ImageSegments.
• Added datasets.SMTP

impute

• Added impute.PreviousImputer.

linear_model

• linear_model.FMClassifier has been moved to the facto module.
• linear_model.FMRegressor has been moved to the facto module.
• Added linear_model.ALMAClassifier.

metrics

• Added metrics.ClassificationReport.
• Added metrics.TimeRolling.
• The implementation of metrics.ROCAUC was incorrect. Using the trapezoidal rule instead of Simpson's rule seems to be more robust.
• metrics.PerClass has been removed; it is recommended that you use metrics.ClassificationReport instead as it gives a better overview.

meta

• Moved meta.TransformedTargetRegressor and meta.BoxCoxRegressor to this module (they were previously in the compose module).
• Added meta.PredClipper

model_selection

• Added model_selection.expand_param_grid to generate a list of models from a grid of parameters.
• Added the model_selection.successive_halving method for selecting hyperparameters.
• The online_score and online_qa_score methods have been merged into a single method named model_selection.progressive_val_score.

preprocessing

• Added preprocessing.RBFSampler.
• Added preprocessing.MaxAbsScaler.
• Added preprocessing.RobustScaler.
• Added preprocessing.Binarizer.
• Added with_mean and with_std parameters to preprocessing.StandardScaler.

optim

• Added optim.losses.BinaryFocalLoss.
• Added the optim.AMSGrad optimizer.
• Added the optim.Nadam optimizer.
• Added optim.losses.Poisson.
• Fixed a performance bug in optim.NesterovMomentum.

reco

• Added reco.FunkMF.
• Renamed reco.SVD to reco.BiasedMF.
• Renamed reco.SGDBaseline to reco.Baseline.
• Models now expect a dict input with user and item fields.

sampling

• Added sampling.RandomUnderSampler.
• Added sampling.RandomOverSampler.
• Added sampling.RandomSampler.
• Added sampling.HardSamplingClassifier.
• Added sampling.HardSamplingRegressor.

stats

• Added stats.AbsMax.
• Added stats.RollingAbsMax.

stream

• Added stream.iter_libsvm.
• stream.iter_csv now supports reading from '.zip' files.
• Added stream.Cache.
• Added a drop parameter to stream.iter_csv to discard fields.