Source code for evalml.pipelines.components.transformers.encoders.target_encoder

import pandas as pd
from woodwork.logical_types import Categorical

from ..transformer import Transformer

from evalml.pipelines.components.transformers.encoders.onehot_encoder import (
    OneHotEncoderMeta
)
from evalml.utils import (
    _convert_woodwork_types_wrapper,
    _retain_custom_types_and_initalize_woodwork,
    import_or_raise,
    infer_feature_types
)


class TargetEncoder(Transformer, metaclass=OneHotEncoderMeta):
    """Target encoder to encode categorical data"""
    name = 'Target Encoder'
    hyperparameter_ranges = {}

    def __init__(self,
                 cols=None,
                 smoothing=1.0,
                 handle_unknown='value',
                 handle_missing='value',
                 random_seed=0,
                 **kwargs):
        """Initializes a transformer that encodes categorical features into target encodings.

        Arguments:
            cols (list): Columns to encode. If None, all string columns will be encoded, otherwise only the columns provided will be encoded.
                Defaults to None
            smoothing (float): The smoothing factor to apply. The larger this value is, the more influence the expected target value has
                on the resulting target encodings. Must be strictly larger than 0. Defaults to 1.0
            handle_unknown (string): Determines how to handle unknown categories for a feature encountered. Options are 'value', 'error', nd 'return_nan'.
                Defaults to 'value', which replaces with the target mean
            handle_missing (string): Determines how to handle missing values encountered during `fit` or `transform`. Options are 'value', 'error', and 'return_nan'.
                Defaults to 'value', which replaces with the target mean
            random_seed (int): Seed for the random number generator. Defaults to 0.
            """

        parameters = {"cols": cols,
                      "smoothing": smoothing,
                      "handle_unknown": handle_unknown,
                      "handle_missing": handle_missing}
        parameters.update(kwargs)

        unknown_and_missing_input_options = ['error', 'return_nan', 'value']
        if handle_unknown not in unknown_and_missing_input_options:
            raise ValueError("Invalid input '{}' for handle_unknown".format(handle_unknown))
        if handle_missing not in unknown_and_missing_input_options:
            raise ValueError("Invalid input '{}' for handle_missing".format(handle_missing))
        if smoothing <= 0:
            raise ValueError("Smoothing value needs to be strictly larger than 0. {} provided".format(smoothing))

        category_encode = import_or_raise('category_encoders', error_msg='category_encoders not installed. Please install using `pip install category_encoders`')
        super().__init__(parameters=parameters,
                         component_obj=category_encode.target_encoder.TargetEncoder(**parameters),
                         random_seed=random_seed)

    def fit(self, X, y):
        return super().fit(X, y)

    def transform(self, X, y=None):
        X_ww = infer_feature_types(X)
        X = _convert_woodwork_types_wrapper(X_ww.to_dataframe())
        if y is not None:
            y = infer_feature_types(y)
            y = _convert_woodwork_types_wrapper(y.to_series())
        X_t = self._component_obj.transform(X, y)
        X_t_df = pd.DataFrame(X_t, columns=X.columns, index=X.index)
        return _retain_custom_types_and_initalize_woodwork(X_ww, X_t_df, ltypes_to_ignore=[Categorical])

    def fit_transform(self, X, y):
        return self.fit(X, y).transform(X, y)

    def get_feature_names(self):
        """Return feature names for the input features after fitting.

        Returns:
            np.array: The feature names after encoding
        """
        return self._component_obj.get_feature_names()

    def _get_feature_provenance(self):
        return {col: col for col in self.get_feature_names()}