|
1
|
|
|
""" |
|
2
|
|
|
Functions for data preprocessing. |
|
3
|
|
|
|
|
4
|
|
|
:author: Andreas Kanz |
|
5
|
|
|
|
|
6
|
|
|
""" |
|
7
|
|
|
|
|
8
|
|
|
# Imports |
|
9
|
1 |
|
import numpy as np |
|
10
|
1 |
|
import pandas as pd |
|
11
|
1 |
|
from sklearn.base import BaseEstimator, TransformerMixin |
|
12
|
1 |
|
from sklearn.ensemble import ExtraTreesRegressor |
|
13
|
1 |
|
from sklearn.experimental import enable_iterative_imputer # noqa |
|
14
|
1 |
|
from sklearn.feature_selection import ( |
|
15
|
|
|
SelectFromModel, |
|
16
|
|
|
SelectPercentile, |
|
17
|
|
|
VarianceThreshold, |
|
18
|
|
|
f_classif, |
|
19
|
|
|
) |
|
20
|
1 |
|
from sklearn.impute import IterativeImputer, SimpleImputer |
|
21
|
1 |
|
from sklearn.linear_model import LassoCV |
|
22
|
1 |
|
from sklearn.model_selection import train_test_split |
|
23
|
1 |
|
from sklearn.pipeline import make_pipeline |
|
24
|
1 |
|
from sklearn.preprocessing import MaxAbsScaler, OneHotEncoder, RobustScaler |
|
25
|
|
|
|
|
26
|
1 |
|
from klib.utils import ( |
|
27
|
|
|
_validate_input_int, |
|
28
|
|
|
_validate_input_range, |
|
29
|
|
|
_validate_input_sum_smaller, |
|
30
|
|
|
) |
|
31
|
|
|
|
|
32
|
1 |
|
__all__ = ["feature_selection_pipe", "num_pipe", "cat_pipe", "train_dev_test_split"] |
|
33
|
|
|
|
|
34
|
|
|
|
|
35
|
1 |
|
class ColumnSelector(BaseEstimator, TransformerMixin): |
|
36
|
|
|
""" |
|
37
|
|
|
Determines and selects numerical and categorical columns from a dataset based on \ |
|
38
|
|
|
their supposed dtype. Unlike sklearn's make_column_selector() missing values are \ |
|
39
|
|
|
temporarily filled in to allow convert_dtypes() to determine the dtype of a column. |
|
40
|
|
|
|
|
41
|
|
|
Parameter |
|
42
|
|
|
--------- |
|
43
|
|
|
num: default, True |
|
44
|
|
|
Select only numerical Columns. If num = False, only categorical columns are \ |
|
45
|
|
|
selected. |
|
46
|
|
|
|
|
47
|
|
|
Returns |
|
48
|
|
|
------- |
|
49
|
|
|
Dataset containing only numerical or categorical data. |
|
50
|
|
|
""" |
|
51
|
|
|
|
|
52
|
1 |
|
def __init__(self, num=True): |
|
53
|
|
|
self.num = num |
|
54
|
|
|
|
|
55
|
1 |
|
def fit(self, X, y=None): |
|
56
|
|
|
return self |
|
57
|
|
|
|
|
58
|
1 |
|
def transform(self, X, y=None): |
|
59
|
|
|
temp = X.fillna(X.mode().iloc[0]).convert_dtypes() |
|
60
|
|
|
|
|
61
|
|
|
if self.num: |
|
62
|
|
|
return X[temp.select_dtypes(include=["number"]).columns.tolist()] |
|
63
|
|
|
return X[temp.select_dtypes(exclude=["number"]).columns.tolist()] |
|
64
|
|
|
|
|
65
|
|
|
|
|
66
|
1 |
|
class PipeInfo(BaseEstimator, TransformerMixin): |
|
67
|
|
|
""" |
|
68
|
|
|
Prints intermediary information about the dataset from within a pipeline. |
|
69
|
|
|
|
|
70
|
|
|
Include at any point in a Pipeline to print out the shape of the dataset at this \ |
|
71
|
|
|
point and to receive an indication of the progress within the pipeline. |
|
72
|
|
|
|
|
73
|
|
|
Set to 'None' to avoid printing the shape of the dataset. This parameter can also \ |
|
74
|
|
|
be set as a hyperparameter, e.g. 'pipeline__pipeinfo-1': [None] or \ |
|
75
|
|
|
'pipeline__pipeinfo-1__name': ['my_custom_name']. |
|
76
|
|
|
|
|
77
|
|
|
Parameter |
|
78
|
|
|
--------- |
|
79
|
|
|
name: string, default None |
|
80
|
|
|
Provide a name for the current step. |
|
81
|
|
|
|
|
82
|
|
|
Returns |
|
83
|
|
|
------- |
|
84
|
|
|
Data: Data is being passed through. |
|
85
|
|
|
""" |
|
86
|
|
|
|
|
87
|
1 |
|
def __init__(self, name=None): |
|
88
|
1 |
|
self.name = name |
|
89
|
|
|
|
|
90
|
1 |
|
def fit(self, X, y=None): |
|
91
|
|
|
return self |
|
92
|
|
|
|
|
93
|
1 |
|
def transform(self, X, y=None): |
|
94
|
|
|
print(f"Step: {self.name} --- Shape: {X.shape}") |
|
95
|
|
|
return X |
|
96
|
|
|
|
|
97
|
|
|
|
|
98
|
1 |
|
def cat_pipe( |
|
99
|
|
|
imputer=SimpleImputer(strategy="most_frequent"), |
|
100
|
|
|
encoder=OneHotEncoder(handle_unknown="ignore"), |
|
101
|
|
|
scaler=MaxAbsScaler(), |
|
102
|
|
|
encoder_info=PipeInfo(name="after encoding categorical data"), |
|
103
|
|
|
): |
|
104
|
|
|
""" |
|
105
|
|
|
Standard preprocessing operations on categorical data. |
|
106
|
|
|
|
|
107
|
|
|
Parameters |
|
108
|
|
|
---------- |
|
109
|
|
|
imputer: default, SimpleImputer(strategy='most_frequent') |
|
110
|
|
|
|
|
111
|
|
|
encoder: default, OneHotEncoder(handle_unknown='ignore') |
|
112
|
|
|
Encode categorical features as a one-hot numeric array. |
|
113
|
|
|
|
|
114
|
|
|
scaler: default, MaxAbsScaler() |
|
115
|
|
|
Scale each feature by its maximum absolute value. MaxAbsScaler() does not \ |
|
116
|
|
|
shift/center the data, and thus does not destroy any sparsity. It is \ |
|
117
|
|
|
recommended to check for outliers before applying MaxAbsScaler(). |
|
118
|
|
|
|
|
119
|
|
|
encoder_info: |
|
120
|
|
|
Prints the shape of the dataset at the end of 'cat_pipe'. Set to 'None' to \ |
|
121
|
|
|
avoid printing the shape of dataset. This parameter can also be set as a \ |
|
122
|
|
|
hyperparameter, e.g. 'pipeline__pipeinfo-1': [None] or \ |
|
123
|
|
|
'pipeline__pipeinfo-1__name': ['my_custom_name']. |
|
124
|
|
|
|
|
125
|
|
|
Returns |
|
126
|
|
|
------- |
|
127
|
|
|
Pipeline |
|
128
|
|
|
""" |
|
129
|
|
|
return make_pipeline( |
|
130
|
|
|
ColumnSelector(num=False), imputer, encoder, encoder_info, scaler |
|
131
|
|
|
) |
|
132
|
|
|
|
|
133
|
|
|
|
|
134
|
1 |
|
def feature_selection_pipe( |
|
135
|
|
|
var_thresh=VarianceThreshold(threshold=0.1), |
|
136
|
|
|
select_from_model=SelectFromModel( |
|
137
|
|
|
LassoCV(cv=4, random_state=408), threshold="0.1*median" |
|
138
|
|
|
), |
|
139
|
|
|
select_percentile=SelectPercentile(f_classif, percentile=95), |
|
140
|
|
|
var_thresh_info=PipeInfo(name="after var_thresh"), |
|
141
|
|
|
select_from_model_info=PipeInfo(name="after select_from_model"), |
|
142
|
|
|
select_percentile_info=PipeInfo(name="after select_percentile"), |
|
143
|
|
|
): |
|
144
|
|
|
""" |
|
145
|
|
|
Preprocessing operations for feature selection. |
|
146
|
|
|
|
|
147
|
|
|
Parameters |
|
148
|
|
|
---------- |
|
149
|
|
|
var_thresh: default, VarianceThreshold(threshold=0.1) |
|
150
|
|
|
Specify a threshold to drop low variance features. |
|
151
|
|
|
|
|
152
|
|
|
select_from_model: default, SelectFromModel(LassoCV(cv=4, random_state=408), \ |
|
153
|
|
|
threshold="0.1 * median") |
|
154
|
|
|
Specify an estimator which is used for selecting features based on importance \ |
|
155
|
|
|
weights. |
|
156
|
|
|
|
|
157
|
|
|
select_percentile: default, SelectPercentile(f_classif, percentile=95) |
|
158
|
|
|
Specify a score-function and a percentile value of features to keep. |
|
159
|
|
|
|
|
160
|
|
|
var_thresh_info, select_from_model_info, select_percentile_info |
|
161
|
|
|
Prints the shape of the dataset after applying the respective function. Set to \ |
|
162
|
|
|
'None' to avoid printing the shape of dataset. This parameter can also be set \ |
|
163
|
|
|
as a hyperparameter, e.g. 'pipeline__pipeinfo-1': [None] \ |
|
164
|
|
|
or 'pipeline__pipeinfo-1__name': ['my_custom_name']. |
|
165
|
|
|
|
|
166
|
|
|
Returns |
|
167
|
|
|
------- |
|
168
|
|
|
Pipeline |
|
169
|
|
|
""" |
|
170
|
|
|
return make_pipeline( |
|
171
|
|
|
var_thresh, |
|
172
|
|
|
var_thresh_info, |
|
173
|
|
|
select_from_model, |
|
174
|
|
|
select_from_model_info, |
|
175
|
|
|
select_percentile, |
|
176
|
|
|
select_percentile_info, |
|
177
|
|
|
) |
|
178
|
|
|
|
|
179
|
|
|
|
|
180
|
1 |
|
def num_pipe( |
|
181
|
|
|
imputer=IterativeImputer( |
|
182
|
|
|
estimator=ExtraTreesRegressor(n_estimators=25, n_jobs=4, random_state=408), |
|
183
|
|
|
random_state=408, |
|
184
|
|
|
), |
|
185
|
|
|
scaler=RobustScaler(), |
|
186
|
|
|
): |
|
187
|
|
|
""" |
|
188
|
|
|
Standard preprocessing operations on numerical data. |
|
189
|
|
|
|
|
190
|
|
|
Parameters |
|
191
|
|
|
---------- |
|
192
|
|
|
imputer: default, IterativeImputer(estimator=ExtraTreesRegressor(n_estimators=25, \ |
|
193
|
|
|
n_jobs=4, random_state=408), random_state=408) |
|
194
|
|
|
|
|
195
|
|
|
scaler: default, RobustScaler() |
|
196
|
|
|
|
|
197
|
|
|
Returns |
|
198
|
|
|
------- |
|
199
|
|
|
Pipeline |
|
200
|
|
|
""" |
|
201
|
|
|
return make_pipeline(ColumnSelector(), imputer, scaler) |
|
202
|
|
|
|
|
203
|
|
|
|
|
204
|
1 |
|
def train_dev_test_split( |
|
205
|
|
|
data, target, dev_size=0.1, test_size=0.1, stratify=None, random_state=408 |
|
206
|
|
|
): |
|
207
|
|
|
""" |
|
208
|
|
|
Split a dataset and a label column into train, dev and test sets. |
|
209
|
|
|
|
|
210
|
|
|
Parameters |
|
211
|
|
|
---------- |
|
212
|
|
|
data: 2D dataset that can be coerced into Pandas DataFrame. If a Pandas DataFrame \ |
|
213
|
|
|
is provided, the index/column information is used to label the plots. |
|
214
|
|
|
|
|
215
|
|
|
target: string, list, np.array or pd.Series, default None |
|
216
|
|
|
Specify target for correlation. E.g. label column to generate only the \ |
|
217
|
|
|
correlations between each feature and the label. |
|
218
|
|
|
|
|
219
|
|
|
dev_size: float, default 0.1 |
|
220
|
|
|
If float, should be between 0.0 and 1.0 and represent the proportion of the \ |
|
221
|
|
|
dataset to include in the dev split. |
|
222
|
|
|
|
|
223
|
|
|
test_size: float, default 0.1 |
|
224
|
|
|
If float, should be between 0.0 and 1.0 and represent the proportion of the \ |
|
225
|
|
|
dataset to include in the test split. |
|
226
|
|
|
|
|
227
|
|
|
stratify: target column, default None |
|
228
|
|
|
If not None, data is split in a stratified fashion, using the input as the \ |
|
229
|
|
|
class labels. |
|
230
|
|
|
|
|
231
|
|
|
random_state: integer, default 408 |
|
232
|
|
|
Random_state is the seed used by the random number generator. |
|
233
|
|
|
|
|
234
|
|
|
Returns |
|
235
|
|
|
------- |
|
236
|
|
|
tuple: Tuple containing train-dev-test split of inputs. |
|
237
|
|
|
""" |
|
238
|
|
|
# Validate Inputs |
|
239
|
1 |
|
_validate_input_range(dev_size, "dev_size", 0, 1) |
|
240
|
1 |
|
_validate_input_range(test_size, "test_size", 0, 1) |
|
241
|
1 |
|
_validate_input_int(random_state, "random_state") |
|
242
|
1 |
|
_validate_input_sum_smaller(1, "Dev and test", dev_size, test_size) |
|
243
|
|
|
|
|
244
|
1 |
|
target_data = [] |
|
245
|
1 |
|
if isinstance(target, str): |
|
246
|
1 |
|
target_data = data[target] |
|
247
|
1 |
|
data = data.drop(target, axis=1) |
|
248
|
|
|
|
|
249
|
1 |
|
elif isinstance(target, (list, pd.Series, np.ndarray)): |
|
250
|
1 |
|
target_data = pd.Series(target) |
|
251
|
|
|
|
|
252
|
1 |
|
X_train, X_dev_test, y_train, y_dev_test = train_test_split( |
|
253
|
|
|
data, |
|
254
|
|
|
target_data, |
|
255
|
|
|
test_size=dev_size + test_size, |
|
256
|
|
|
random_state=random_state, |
|
257
|
|
|
stratify=stratify, |
|
258
|
|
|
) |
|
259
|
|
|
|
|
260
|
1 |
|
if (dev_size == 0) or (test_size == 0): |
|
261
|
1 |
|
return X_train, X_dev_test, y_train, y_dev_test |
|
262
|
1 |
|
X_dev, X_test, y_dev, y_test = train_test_split( |
|
263
|
|
|
X_dev_test, |
|
264
|
|
|
y_dev_test, |
|
265
|
|
|
test_size=test_size / (dev_size + test_size), |
|
266
|
|
|
random_state=random_state, |
|
267
|
|
|
stratify=y_dev_test, |
|
268
|
|
|
) |
|
269
|
|
|
return X_train, X_dev, X_test, y_train, y_dev, y_test |
|
270
|
|
|
|
akanz1 /
klib
