0.17.0 - 2023-05-27

bandit

• Bandit policies now return a single arm when the pull method is called, instead of yielding or one more arms at a time. This is simpler to understand. We will move back to multi-armed pulls in the future.
• Added bandit.Exp3.
• bandit.UCB and bandit.Exp3 have an extra reward_scaler parameter, which can be any object that inherits from compose.TargetTransformRegressor. This allows scaling rewards before updating arms.

compose

• compose.TransformerProduct now correctly returns a compose.TransformerUnion when a transformer is added to it.
• Fixed compose.TransformerProduct's transform_many behavior.
• compose.TransformerUnion and compose.TransformerProduct will now clone the provided estimators, so that shallow copies aren't shared in different places.

model_selection

• Added model_selection.BanditClassifier, which is the classification equivalent to bandit.BanditRegressor. Both are methods to perform online model selection via a bandit policy.

multioutput

• metrics.multioutput.MacroAverage and metrics.multioutput.MicroAverage now loop over the keys of y_true instead of y_pred. This ensures a KeyError is correctly raised if y_pred is missing an output that is present in y_true.

preprocessing

• Added preprocessing.TargetMinMaxScaler, which operates the same as preprocessing.TargetStandardScaler, but instead uses min-max scaling.