"""Target Encoder"""
import numpy as np
import pandas as pd
from scipy.special import expit
from category_encoders.ordinal import OrdinalEncoder
import category_encoders.utils as util
__author__ = 'chappers'
[docs]class TargetEncoder(util.BaseEncoder, util.SupervisedTransformerMixin):
"""Target encoding for categorical features.
Supported targets: binomial and continuous. For polynomial target support, see PolynomialWrapper.
For the case of categorical target: features are replaced with a blend of posterior probability of the target
given particular categorical value and the prior probability of the target over all the training data.
For the case of continuous target: features are replaced with a blend of the expected value of the target
given particular categorical value and the expected value of the target over all the training data.
Parameters
----------
verbose: int
integer indicating verbosity of the output. 0 for none.
cols: list
a list of columns to encode, if None, all string columns will be encoded.
drop_invariant: bool
boolean for whether or not to drop columns with 0 variance.
return_df: bool
boolean for whether to return a pandas DataFrame from transform (otherwise it will be a numpy array).
handle_missing: str
options are 'error', 'return_nan' and 'value', defaults to 'value', which returns the target mean.
handle_unknown: str
options are 'error', 'return_nan' and 'value', defaults to 'value', which returns the target mean.
min_samples_leaf: int
For regularization the weighted average between category mean and global mean is taken. The weight is
an S-shaped curve between 0 and 1 with the number of samples for a category on the x-axis.
The curve reaches 0.5 at min_samples_leaf. (parameter k in the original paper)
smoothing: float
smoothing effect to balance categorical average vs prior. Higher value means stronger regularization.
The value must be strictly bigger than 0. Higher values mean a flatter S-curve (see min_samples_leaf).
hierarchy: dict or dataframe
A dictionary or a dataframe to define the hierarchy for mapping.
If a dictionary, this contains a dict of columns to map into hierarchies. Dictionary key(s) should be the column name from X
which requires mapping. For multiple hierarchical maps, this should be a dictionary of dictionaries.
If dataframe: a dataframe defining columns to be used for the hierarchies. Column names must take the form:
HIER_colA_1, ... HIER_colA_N, HIER_colB_1, ... HIER_colB_M, ...
where [colA, colB, ...] are given columns in cols list.
1:N and 1:M define the hierarchy for each column where 1 is the highest hierarchy (top of the tree). A single column or multiple
can be used, as relevant.
Examples
-------
>>> from category_encoders import *
>>> import pandas as pd
>>> from sklearn.datasets import fetch_openml
>>> display_cols = ["Id", "MSSubClass", "MSZoning", "LotFrontage", "YearBuilt", "Heating", "CentralAir"]
>>> bunch = fetch_openml(name="house_prices", as_frame=True)
>>> y = bunch.target > 200000
>>> X = pd.DataFrame(bunch.data, columns=bunch.feature_names)[display_cols]
>>> enc = TargetEncoder(cols=['CentralAir', 'Heating'], min_samples_leaf=20, smoothing=10).fit(X, y)
>>> numeric_dataset = enc.transform(X)
>>> print(numeric_dataset.info())
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 1460 entries, 0 to 1459
Data columns (total 7 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 Id 1460 non-null float64
1 MSSubClass 1460 non-null float64
2 MSZoning 1460 non-null object
3 LotFrontage 1201 non-null float64
4 YearBuilt 1460 non-null float64
5 Heating 1460 non-null float64
6 CentralAir 1460 non-null float64
dtypes: float64(6), object(1)
memory usage: 80.0+ KB
None
>>> from category_encoders.datasets import load_compass
>>> X, y = load_compass()
>>> hierarchical_map = {'compass': {'N': ('N', 'NE'), 'S': ('S', 'SE'), 'W': 'W'}}
>>> enc = TargetEncoder(verbose=1, smoothing=2, min_samples_leaf=2, hierarchy=hierarchical_map, cols=['compass']).fit(X.loc[:,['compass']], y)
>>> hierarchy_dataset = enc.transform(X.loc[:,['compass']])
>>> print(hierarchy_dataset['compass'].values)
[0.62263617 0.62263617 0.90382995 0.90382995 0.90382995 0.17660024
0.17660024 0.46051953 0.46051953 0.46051953 0.46051953 0.40332791
0.40332791 0.40332791 0.40332791 0.40332791]
>>> X, y = load_postcodes('binary')
>>> cols = ['postcode']
>>> HIER_cols = ['HIER_postcode_1','HIER_postcode_2','HIER_postcode_3','HIER_postcode_4']
>>> enc = TargetEncoder(verbose=1, smoothing=2, min_samples_leaf=2, hierarchy=X[HIER_cols], cols=['postcode']).fit(X['postcode'], y)
>>> hierarchy_dataset = enc.transform(X['postcode'])
>>> print(hierarchy_dataset.loc[0:10, 'postcode'].values)
[0.75063473 0.90208756 0.88328833 0.77041254 0.68891504 0.85012847
0.76772574 0.88742357 0.7933824 0.63776756 0.9019973 ]
References
----------
.. [1] A Preprocessing Scheme for High-Cardinality Categorical Attributes in Classification and Prediction Problems, from
https://dl.acm.org/citation.cfm?id=507538
"""
prefit_ordinal = True
encoding_relation = util.EncodingRelation.ONE_TO_ONE
def __init__(self, verbose=0, cols=None, drop_invariant=False, return_df=True, handle_missing='value',
handle_unknown='value', min_samples_leaf=20, smoothing=10, hierarchy=None):
super().__init__(verbose=verbose, cols=cols, drop_invariant=drop_invariant, return_df=return_df,
handle_unknown=handle_unknown, handle_missing=handle_missing)
self.ordinal_encoder = None
self.min_samples_leaf = min_samples_leaf
self.smoothing = smoothing
self.mapping = None
self._mean = None
if isinstance(hierarchy, (dict, pd.DataFrame)) and cols is None:
raise ValueError('Hierarchy is defined but no columns are named for encoding')
if isinstance(hierarchy, dict):
self.hierarchy = {}
self.hierarchy_depth = {}
for switch in hierarchy:
flattened_hierarchy = util.flatten_reverse_dict(hierarchy[switch])
hierarchy_check = self._check_dict_key_tuples(flattened_hierarchy)
self.hierarchy_depth[switch] = hierarchy_check[1]
if not hierarchy_check[0]:
raise ValueError('Hierarchy mapping contains different levels for key "' + switch + '"')
self.hierarchy[switch] = {(k if isinstance(t, tuple) else t): v for t, v in flattened_hierarchy.items() for k in t}
elif isinstance(hierarchy, pd.DataFrame):
self.hierarchy = hierarchy
self.hierarchy_depth = {}
for col in self.cols:
HIER_cols = self.hierarchy.columns[self.hierarchy.columns.str.startswith(f'HIER_{col}')].values
HIER_levels = [int(i.replace(f'HIER_{col}_', '')) for i in HIER_cols]
if np.array_equal(sorted(HIER_levels), np.arange(1, max(HIER_levels)+1)):
self.hierarchy_depth[col] = max(HIER_levels)
else:
raise ValueError(f'Hierarchy columns are not complete for column {col}')
elif hierarchy is None:
self.hierarchy = hierarchy
else:
raise ValueError('Given hierarchy mapping is neither a dictionary nor a dataframe')
self.cols_hier = []
def _check_dict_key_tuples(self, d):
min_tuple_size = min(len(v) for v in d.values())
max_tuple_size = max(len(v) for v in d.values())
return min_tuple_size == max_tuple_size, min_tuple_size
def _fit(self, X, y, **kwargs):
if isinstance(self.hierarchy, dict):
X_hier = pd.DataFrame()
for switch in self.hierarchy:
if switch in self.cols:
colnames = [f'HIER_{str(switch)}_{str(i + 1)}' for i in range(self.hierarchy_depth[switch])]
df = pd.DataFrame(X[str(switch)].map(self.hierarchy[str(switch)]).tolist(), index=X.index, columns=colnames)
X_hier = pd.concat([X_hier, df], axis=1)
elif isinstance(self.hierarchy, pd.DataFrame):
X_hier = self.hierarchy
if isinstance(self.hierarchy, (dict, pd.DataFrame)):
enc_hier = OrdinalEncoder(
verbose=self.verbose,
cols=X_hier.columns,
handle_unknown='value',
handle_missing='value'
)
enc_hier = enc_hier.fit(X_hier)
X_hier_ordinal = enc_hier.transform(X_hier)
self.ordinal_encoder = OrdinalEncoder(
verbose=self.verbose,
cols=self.cols,
handle_unknown='value',
handle_missing='value'
)
self.ordinal_encoder = self.ordinal_encoder.fit(X)
X_ordinal = self.ordinal_encoder.transform(X)
if self.hierarchy is not None:
self.mapping = self.fit_target_encoding(pd.concat([X_ordinal, X_hier_ordinal], axis=1), y)
else:
self.mapping = self.fit_target_encoding(X_ordinal, y)
def fit_target_encoding(self, X, y):
mapping = {}
prior = self._mean = y.mean()
for switch in self.ordinal_encoder.category_mapping:
col = switch.get('col')
if 'HIER_' not in str(col):
values = switch.get('mapping')
scalar = prior
if (isinstance(self.hierarchy, dict) and col in self.hierarchy) or \
(isinstance(self.hierarchy, pd.DataFrame)):
for i in range(self.hierarchy_depth[col]):
col_hier = 'HIER_'+str(col)+'_'+str(i+1)
col_hier_m1 = col if i == self.hierarchy_depth[col]-1 else 'HIER_'+str(col)+'_'+str(i+2)
if not X[col].equals(X[col_hier]) and len(X[col_hier].unique())>1:
stats_hier = y.groupby(X[col_hier]).agg(['count', 'mean'])
smoove_hier = self._weighting(stats_hier['count'])
scalar_hier = scalar * (1 - smoove_hier) + stats_hier['mean'] * smoove_hier
scalar_hier_long = X[[col_hier_m1, col_hier]].drop_duplicates()
scalar_hier_long.index = np.arange(1, scalar_hier_long.shape[0]+1)
scalar = scalar_hier_long[col_hier].map(scalar_hier.to_dict())
stats = y.groupby(X[col]).agg(['count', 'mean'])
smoove = self._weighting(stats['count'])
smoothing = scalar * (1 - smoove) + stats['mean'] * smoove
if self.handle_unknown == 'return_nan':
smoothing.loc[-1] = np.nan
elif self.handle_unknown == 'value':
smoothing.loc[-1] = prior
if self.handle_missing == 'return_nan':
smoothing.loc[values.loc[np.nan]] = np.nan
elif self.handle_missing == 'value':
smoothing.loc[-2] = prior
mapping[col] = smoothing
return mapping
def _transform(self, X, y=None):
# Now X is the correct dimensions it works with pre fitted ordinal encoder
X = self.ordinal_encoder.transform(X)
if self.handle_unknown == 'error':
if X[self.cols].isin([-1]).any().any():
raise ValueError('Unexpected categories found in dataframe')
X = self.target_encode(X)
return X
def target_encode(self, X_in):
X = X_in.copy(deep=True)
# Was not mapping extra columns as self.featuer_names_in did not include new column
for col in self.cols:
X[col] = X[col].map(self.mapping[col])
return X
def _weighting(self, n):
# monotonically increasing function of n bounded between 0 and 1
# sigmoid in this case, using scipy.expit for numerical stability
return expit((n - self.min_samples_leaf) / self.smoothing)