8000 Merge pull request #2027 from mblondel/select_categorical · scikit-learn/scikit-learn@4fe51ba · GitHub
[go: up one dir, main page]
Skip to content

Commit 4fe51ba

Browse files
larsmanslarsmans
committed
Merge pull request #2027 from mblondel/select_categorical
ENH Non-categorical variables in OneHotEncoder
2 parents d420aaf + e11873f commit 4fe51ba

File tree

3 files changed

+165
-28
lines changed

3 files changed

+165
-28
lines changed

‎doc/modules/preprocessing.rst

Lines changed: 4 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -333,15 +333,16 @@ Continuing the example above::
333333

334334
>>> enc = preprocessing.OneHotEncoder()
335335
>>> enc.fit([[0, 0, 3], [1, 1, 0], [0, 2, 1], [1, 0, 2]])
336-
OneHotEncoder(dtype=<type 'float'>, n_values='auto')
336+
OneHotEncoder(categorical_features='all', dtype=<type 'float'>,
337+
n_values='auto')
337338
>>> enc.transform([[0, 1, 3]]).toarray()
338339
array([[ 1., 0., 0., 1., 0., 0., 0., 0., 1.]])
339340

340341
By default, how many values each feature can take is inferred automatically from the dataset.
341-
It is possible to specify this explicitly using the parameter ``n_values``.
342+
It is possible to specify this explicitly using the parameter ``n_values``.
342343
There are two genders, three possible continents and four web browsers in our
343344
dataset.
344-
Then we fit the estimator, and transform a data point.
345+
Then we fit the estimator, and transform a data point.
345346
In the result, the first two numbers encode the gender, the next set of three
346347
numbers the continent and the last four the web browser.
347348

‎sklearn/preprocessing.py

Lines changed: 99 additions & 25 deletions
Original file line numberDiff line numberDiff line change
@@ -12,7 +12,11 @@
1212

1313
from .base import BaseEstimator, TransformerMixin
1414
from .externals.six import string_types
15-
from .utils import check_arrays, array2d, atleast2d_or_csr, safe_asarray
15+
from .utils import check_arrays
16+
from .utils import array2d
17+
from .utils import atleast2d_or_csr
18+
from .utils import atleast2d_or_csc
19+
from .utils import safe_asarray
1620
from .utils import warn_if_not_float
1721
from .utils.fixes import unique
1822

@@ -35,6 +39,7 @@
3539
'LabelEncoder',
3640
'MinMaxScaler',
3741
'Normalizer',
42+
'OneHotEncoder',
3843
'StandardScaler',
3944
'binarize',
4045
'normalize',
@@ -632,6 +637,53 @@ def transform(self, X, y=None, copy=None):
632637
return binarize(X, threshold=self.threshold, copy=copy)
633638

634639

640+
def _transform_selected(X, transform, selected="all"):
641+
"""Apply a transform function to portion of selected features
642+
643+
Parameters
644+
----------
645+
X : array-like or sparse matrix, shape=(n_samples, n_features)
646+
Dense array or sparse matrix.
647+
648+
transform : callable
649+
A callable transform(X) -> X_transformed
650+
651+
selected: "all" or array of indices or mask
652+
Specify what features to apply the transform to.
653+
654+
Returns
655+
-------
656+
X : array or sparse matrix, shape=(n_samples, n_features_new)
657+
"""
658+
if selected == "all":
659+
return transform(X)
660+
elif len(selected) == 0:
661+
return X
662+
else:
663+
X = atleast2d_or_csc(X)
664+
n_features = X.shape[1]
665+
ind = np.arange(n_features)
666+
sel = np.zeros(n_features, dtype=bool)
667+
sel[np.array(selected)] = True
668+
not_sel = np.logical_not(sel)
669+
n_selected = np.sum(sel)
670+
671+
if n_selected == 0:
672+
# No features selected.
673+
return X
674+
elif n_selected == n_features:
675+
# All features selected.
676+
return transform(X)
677+
else:
678+
X_sel = transform(X[:, ind[sel]])
679+
X_not_sel = X[:, ind[not_sel]]
680+
681+
if sp.issparse(X_sel) or sp.issparse(X_not_sel):
682+
return sp.hstack((X_sel, X_not_sel))
683+
else:
684+
return np.hstack((X_sel, X_not_sel))
685+
686+
635687
class OneHotEncoder(BaseEstimator, TransformerMixin):
636688
"""Encode categorical integer features using a one-hot aka one-of-K scheme.
637689
@@ -646,11 +698,21 @@ class OneHotEncoder(BaseEstimator, TransformerMixin):
646698
647699
Parameters
648700
----------
649-
n_values : 'auto', int or array of int
701+
n_values : 'auto', int or array of ints
650702
Number of values per feature.
651-
'auto' : determine value range from training data.
652-
int : maximum value for all features.
653-
array : maximum value per feature.
703+
704+
- 'auto' : determine value range from training data.
705+
- int : maximum value for all features.
706+
- array : maximum value per feature.
707+
708+
categorical_features: "all" or array of indices or mask
709+
Specify what features are treated as categorical.
710+
711+
- 'all' (default): All features are treated as categorical.
712+
- array of indices: Array of categorical feature indices.
713+
- mask: Array of length n_features and with dtype=bool.
714+
715+
Non-categorical features are always stacked to the right of the matrix.
654716
655717
dtype : number type, default=np.float
656718
Desired dtype of output.
@@ -680,7 +742,8 @@ class OneHotEncoder(BaseEstimator, TransformerMixin):
680742
>>> enc = OneHotEncoder()
681743
>>> enc.fit([[0, 0, 3], [1, 1, 0], [0, 2, 1], 10000 \
682744
[1, 0, 2]]) # doctest: +ELLIPSIS
683-
OneHotEncoder(dtype=<... 'float'>, n_values='auto')
745+
OneHotEncoder(categorical_features='all', dtype=<type 'float'>,
746+
n_values='auto')
684747
>>> enc.n_values_
685748
array([2, 3, 4])
686749
>>> enc.feature_indices_
@@ -690,12 +753,13 @@ class OneHotEncoder(BaseEstimator, TransformerMixin):
690753
691754
See also
692755
--------
693-
LabelEncoder : performs a one-hot encoding on arbitrary class labels.
694756
sklearn.feature_extraction.DictVectorizer : performs a one-hot encoding of
695757
dictionary items (also handles string-valued features).
696758
"""
697-
def __init__(self, n_values="auto", dtype=np.float):
759+
def __init__(self, n_values="auto", categorical_features="all",
760+
dtype=np.float):
698761
self.n_values = n_values
762+
self.categorical_features = categorical_features
699763
self.dtype = dtype
700764

701765
def fit(self, X, y=None):
@@ -713,12 +777,8 @@ def fit(self, X, y=None):
713777
self.fit_transform(X)
714778
return self
715779

716-
def fit_transform(self, X, y=None):
717-
"""Fit OneHotEncoder to X, then transform X.
718-
719-
Equivalent to self.fit(X).transform(X), but more convenient and more
720-
efficient. See fit for the parameters, transform for the return value.
721-
"""
780+
def _fit_transform(self, X):
781+
"""Asssumes X contains only categorical features."""
722782
X = check_arrays(X, sparse_format='dense', dtype=np.int)[0]
723783
if np.any(X < 0):
724784
raise ValueError("X needs to contain only non-negative integers.")
@@ -759,19 +819,17 @@ def fit_transform(self, X, y=None):
759819

760820
return out
761821

762-
def transform(self, X):
763-
"""Transform X using one-hot encoding.
764-
765-
Parameters
766-
----------
767-
X : array-like, shape=(n_samples, feature_indices_[-1])
768-
Input array of type int.
822+
def fit_transform(self, X, y=None):
823+
"""Fit OneHotEncoder to X, then transform X.
769824
770-
Returns
771-
-------
772-
X_out : sparse matrix, dtype=int
773-
Transformed input.
825+
Equivalent to self.fit(X).transform(X), but more convenient and more
826+
efficient. See fit for the parameters, transform for the return value.
774827
"""
828+
return _transform_selected(X, self._fit_transform,
829+
self.categorical_features)
830+
831+
def _transform(self, X):
832+
"""Asssumes X contains only categorical features."""
775833
X = check_arrays(X, sparse_format='dense', dtype=np.int)[0]
776834
if np.any(X < 0):
777835
raise ValueError("X needs to contain only non-negative integers.")
@@ -798,6 +856,22 @@ def transform(self, X):
798856
out = out[:, self.active_features_]
799857
return out
800858

859+
def transform(self, X):
860+
"""Transform X using one-hot encoding.
861+
862+
Parameters
863+
----------
< F438 code>864+
X : array-like, shape=(n_samples, n_features)
865+
Input array of type int.
866+
867+
Returns
868+
-------
869+
X_out : sparse matrix, dtype=int
870+
Transformed input.
871+
"""
872+
return _transform_selected(X, self._transform,
873+
self.categorical_features)
874+
801875

802876
class LabelEncoder(BaseEstimator, TransformerMixin):
803877
"""Encode labels with value between 0 and n_classes-1.

‎sklearn/tests/test_preprocessing.py

Lines changed: 62 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -15,6 +15,7 @@
1515
from sklearn.preprocessing import Binarizer
1616
from sklearn.preprocessing import KernelCenterer
1717
from sklearn.preprocessing import LabelBinarizer
18+
from sklearn.preprocessing import _transform_selected
1819
from sklearn.preprocessing import OneHotEncoder
1920
from sklearn.preprocessing import LabelEncoder
2021
from sklearn.preprocessing import Normalizer
@@ -612,6 +613,67 @@ def test_one_hot_encoder():
612613
assert_raises(ValueError, enc.transform, [[0], [-1]])
613614

614615

616+
def _check_transform_selected(X, Xexpected, sel):
617+
for M in (X, sp.csr_matrix(X)):
618+
Xtr = _transform_selected(M, Binarizer().transform, sel)
619+
assert_array_equal(toarray(Xtr), Xexpected)
620+
621+
622+
def test_transform_selected():
623+
X = [[3, 2, 1], [0, 1, 1]]
624+
625+
Xexpected = [[1, 2, 1], [0, 1, 1]]
626+
_check_transform_selected(X, Xexpected, [0])
627+
_check_transform_selected(X, Xexpected, [True, False, False])
628+
629+
Xexpected = [[1, 1, 1], [0, 1, 1]]
630+
_check_transform_selected(X, Xexpected, [0, 1, 2])
631+
_check_transform_selected(X, Xexpected, [True, True, True])
632+
_check_transform_selected(X, Xexpected, "all")
633+
634+
_check_transform_selected(X, X, [])
635+
_check_transform_selected(X, X, [False, False, False])
636+
637+
638+
def _run_one_hot(X, X2, cat):
639+
enc = OneHotEncoder(categorical_features=cat)
640+
Xtr = enc.fit_transform(X)
641+
X2tr = enc.transform(X2)
642+
return Xtr, X2tr
643+
644+
645+
def _check_one_hot(X, X2, cat, n_features):
646+
ind = np.where(cat)[0]
647+
# With mask
648+
A, B = _run_one_hot(X, X2, cat)
649+
# With indices
650+
C, D = _run_one_hot(X, X2, ind)
651+
# Check shape
652+
assert_equal(A.shape, (2, n_features))
653+
assert_equal(B.shape, (1, n_features))
654+
assert_equal(C.shape, (2, n_features))
655+
assert_equal(D.shape, (1, n_features))
656+
# Check that mask and indices give the same results
657+
assert_array_equal(toarray(A), toarray(C))
658+
assert_array_equal(toarray(B), toarray(D))
659+
660+
661+
def test_one_hot_encoder_categorical_features():
662+
X = np.array([[3, 2, 1], [0, 1, 1]])
663+
X2 = np.array([[1, 1, 1]])
664+
665+
cat = [True, False, False]
666+
_check_one_hot(X, X2, cat, 4)
667+
668+
# Edge case: all non-categorical
669+
cat = [False, False, False]
670+
_check_one_hot(X, X2, cat, 3)
671+
672+
# Edge case: all categorical
673+
cat = [True, True, True]
674+
_check_one_hot(X, X2, cat, 5)
675+
676+
615677
def test_label_encoder():
616678
"""Test LabelEncoder's transform and inverse_transform methods"""
617679
le = LabelEncoder()

0 commit comments

Comments
 (0)
0