|
1
|
|
|
''' |
|
2
|
|
|
Functions for data cleaning. |
|
3
|
|
|
|
|
4
|
|
|
:author: Andreas Kanz |
|
5
|
|
|
|
|
6
|
|
|
''' |
|
7
|
|
|
|
|
8
|
|
|
# Imports |
|
9
|
|
|
import itertools |
|
10
|
|
|
import numpy as np |
|
11
|
|
|
import pandas as pd |
|
12
|
|
|
from sklearn.base import BaseEstimator, TransformerMixin |
|
13
|
|
|
|
|
14
|
|
|
from .describe import corr_mat |
|
15
|
|
|
from .utils import (_diff_report, |
|
16
|
|
|
_drop_duplicates, |
|
17
|
|
|
_missing_vals, |
|
18
|
|
|
_validate_input_bool, |
|
19
|
|
|
_validate_input_range) |
|
20
|
|
|
|
|
21
|
|
|
|
|
22
|
|
|
__all__ = ['convert_datatypes', |
|
23
|
|
|
'data_cleaning', |
|
24
|
|
|
'drop_missing', |
|
25
|
|
|
'mv_col_handling'] |
|
26
|
|
|
|
|
27
|
|
|
|
|
28
|
|
|
def optimize_ints(data): |
|
29
|
|
|
data = pd.DataFrame(data).copy() |
|
30
|
|
|
ints = data.select_dtypes(include=['int64']).columns.tolist() |
|
31
|
|
|
data[ints] = data[ints].apply(pd.to_numeric, downcast='integer') |
|
32
|
|
|
return data |
|
33
|
|
|
|
|
34
|
|
|
|
|
35
|
|
|
def optimize_floats(data): |
|
36
|
|
|
data = pd.DataFrame(data).copy() |
|
37
|
|
|
floats = data.select_dtypes(include=['float64']).columns.tolist() |
|
38
|
|
|
data[floats] = data[floats].apply(pd.to_numeric, downcast='float') |
|
39
|
|
|
return data |
|
40
|
|
|
|
|
41
|
|
|
|
|
42
|
|
|
def convert_datatypes(data, category=True, cat_threshold=0.05, cat_exclude=None): |
|
43
|
|
|
''' |
|
44
|
|
|
Converts columns to best possible dtypes using dtypes supporting pd.NA. Temporarily not converting to integers \ |
|
45
|
|
|
due to an issue in pandas. This is expected to be fixed in pandas 1.1. \ |
|
46
|
|
|
See https://github.com/pandas-dev/pandas/issues/33803 |
|
47
|
|
|
|
|
48
|
|
|
Parameters |
|
49
|
|
|
---------- |
|
50
|
|
|
data: 2D dataset that can be coerced into Pandas DataFrame. |
|
51
|
|
|
|
|
52
|
|
|
category: bool, default True |
|
53
|
|
|
Change dtypes of columns with dtype "object" to "category". Set threshold using cat_threshold or exclude \ |
|
54
|
|
|
columns using cat_exclude. |
|
55
|
|
|
|
|
56
|
|
|
cat_threshold: float, default 0.05 |
|
57
|
|
|
Ratio of unique values below which categories are inferred and column dtype is changed to categorical. |
|
58
|
|
|
|
|
59
|
|
|
cat_exclude: list, default None |
|
60
|
|
|
List of columns to exclude from categorical conversion. |
|
61
|
|
|
|
|
62
|
|
|
Returns |
|
63
|
|
|
------- |
|
64
|
|
|
data: Pandas DataFrame |
|
65
|
|
|
''' |
|
66
|
|
|
|
|
67
|
|
|
# Validate Inputs |
|
68
|
|
|
_validate_input_bool(category, 'Category') |
|
69
|
|
|
_validate_input_range(cat_threshold, 'cat_threshold', 0, 1) |
|
70
|
|
|
|
|
71
|
|
|
cat_exclude = [] if cat_exclude is None else cat_exclude.copy() |
|
72
|
|
|
|
|
73
|
|
|
data = pd.DataFrame(data).copy() |
|
74
|
|
|
for col in data.columns: |
|
75
|
|
|
unique_vals_ratio = data[col].nunique(dropna=False) / data.shape[0] |
|
76
|
|
|
if (category and |
|
77
|
|
|
unique_vals_ratio < cat_threshold and |
|
78
|
|
|
col not in cat_exclude and |
|
79
|
|
|
data[col].dtype == 'object'): |
|
80
|
|
|
data[col] = data[col].astype('category') |
|
81
|
|
|
data[col] = data[col].convert_dtypes(infer_objects=True, convert_string=True, |
|
82
|
|
|
convert_integer=False, convert_boolean=True) |
|
83
|
|
|
|
|
84
|
|
|
data = optimize_ints(data) |
|
85
|
|
|
data = optimize_floats(data) |
|
86
|
|
|
|
|
87
|
|
|
return data |
|
88
|
|
|
|
|
89
|
|
|
|
|
90
|
|
|
def drop_missing(data, drop_threshold_cols=1, drop_threshold_rows=1): |
|
91
|
|
|
''' |
|
92
|
|
|
Drops completely empty columns and rows by default and optionally provides flexibility to loosen restrictions to \ |
|
93
|
|
|
drop additional columns and rows based on the fraction of remaining NA-values. |
|
94
|
|
|
|
|
95
|
|
|
Parameters |
|
96
|
|
|
---------- |
|
97
|
|
|
data: 2D dataset that can be coerced into Pandas DataFrame. |
|
98
|
|
|
|
|
99
|
|
|
drop_threshold_cols: float, default 1 |
|
100
|
|
|
Drop columns with NA-ratio above the specified threshold. |
|
101
|
|
|
|
|
102
|
|
|
drop_threshold_rows: float, default 1 |
|
103
|
|
|
Drop rows with NA-ratio above the specified threshold. |
|
104
|
|
|
|
|
105
|
|
|
Returns |
|
106
|
|
|
------- |
|
107
|
|
|
data_cleaned: Pandas DataFrame |
|
108
|
|
|
|
|
109
|
|
|
Notes |
|
110
|
|
|
----- |
|
111
|
|
|
Columns are dropped first. Rows are dropped based on the remaining data. |
|
112
|
|
|
''' |
|
113
|
|
|
|
|
114
|
|
|
# Validate Inputs |
|
115
|
|
|
_validate_input_range(drop_threshold_cols, 'drop_threshold_cols', 0, 1) |
|
116
|
|
|
_validate_input_range(drop_threshold_rows, 'drop_threshold_rows', 0, 1) |
|
117
|
|
|
|
|
118
|
|
|
data = pd.DataFrame(data).copy() |
|
119
|
|
|
data = data.dropna(axis=0, how='all').dropna(axis=1, how='all') |
|
120
|
|
|
data = data.drop(columns=data.loc[:, _missing_vals(data)['mv_cols_ratio'] > drop_threshold_cols].columns) |
|
121
|
|
|
data_cleaned = data.drop(index=data.loc[_missing_vals(data)['mv_rows_ratio'] > drop_threshold_rows, :].index) |
|
122
|
|
|
|
|
123
|
|
|
return data_cleaned |
|
124
|
|
|
|
|
125
|
|
|
|
|
126
|
|
|
def data_cleaning(data, drop_threshold_cols=0.9, drop_threshold_rows=0.9, drop_duplicates=True, |
|
127
|
|
|
convert_dtypes=True, category=True, cat_threshold=0.03, cat_exclude=None, show='changes'): |
|
128
|
|
|
''' |
|
129
|
|
|
Perform initial data cleaning tasks on a dataset, such as dropping single valued and empty rows, empty \ |
|
130
|
|
|
columns as well as optimizing the datatypes. |
|
131
|
|
|
|
|
132
|
|
|
Parameters |
|
133
|
|
|
---------- |
|
134
|
|
|
data: 2D dataset that can be coerced into Pandas DataFrame. |
|
135
|
|
|
|
|
136
|
|
|
drop_threshold_cols: float, default 0.9 |
|
137
|
|
|
Drop columns with NA-ratio above the specified threshold. |
|
138
|
|
|
|
|
139
|
|
|
drop_threshold_rows: float, default 0.9 |
|
140
|
|
|
Drop rows with NA-ratio above the specified threshold. |
|
141
|
|
|
|
|
142
|
|
|
drop_duplicates: bool, default True |
|
143
|
|
|
Drop duplicate rows, keeping the first occurence. This step comes after the dropping of missing values. |
|
144
|
|
|
|
|
145
|
|
|
convert_dtypes: bool, default True |
|
146
|
|
|
Convert dtypes using pd.convert_dtypes(). |
|
147
|
|
|
|
|
148
|
|
|
category: bool, default True |
|
149
|
|
|
Enable changing dtypes of 'object' columns to "category". Set threshold using cat_threshold. Requires \ |
|
150
|
|
|
convert_dtypes=True. |
|
151
|
|
|
|
|
152
|
|
|
cat_threshold: float, default 0.03 |
|
153
|
|
|
Ratio of unique values below which categories are inferred and column dtype is changed to categorical. |
|
154
|
|
|
|
|
155
|
|
|
cat_exclude: list, default None |
|
156
|
|
|
List of columns to exclude from categorical conversion. |
|
157
|
|
|
|
|
158
|
|
|
show: {'all', 'changes', None} default 'all' |
|
159
|
|
|
Specify verbosity of the output. |
|
160
|
|
|
* 'all': Print information about the data before and after cleaning as well as information about changes. |
|
161
|
|
|
* 'changes': Print out differences in the data before and after cleaning. |
|
162
|
|
|
* None: No information about the data and the data cleaning is printed. |
|
163
|
|
|
|
|
164
|
|
|
Returns |
|
165
|
|
|
------- |
|
166
|
|
|
data_cleaned: Pandas DataFrame |
|
167
|
|
|
|
|
168
|
|
|
See Also |
|
169
|
|
|
-------- |
|
170
|
|
|
convert_datatypes: Convert columns to best possible dtypes. |
|
171
|
|
|
drop_missing : Flexibly drop columns and rows. |
|
172
|
|
|
_memory_usage: Gives the total memory usage in kilobytes. |
|
173
|
|
|
_missing_vals: Metrics about missing values in the dataset. |
|
174
|
|
|
|
|
175
|
|
|
Notes |
|
176
|
|
|
----- |
|
177
|
|
|
The category dtype is not grouped in the summary, unless it contains exactly the same categories. |
|
178
|
|
|
''' |
|
179
|
|
|
|
|
180
|
|
|
# Validate Inputs |
|
181
|
|
|
_validate_input_range(drop_threshold_cols, 'drop_threshold_cols', 0, 1) |
|
182
|
|
|
_validate_input_range(drop_threshold_rows, 'drop_threshold_rows', 0, 1) |
|
183
|
|
|
_validate_input_bool(drop_duplicates, 'drop_duplicates') |
|
184
|
|
|
_validate_input_bool(convert_dtypes, 'convert_datatypes') |
|
185
|
|
|
_validate_input_bool(category, 'category') |
|
186
|
|
|
_validate_input_range(cat_threshold, 'cat_threshold', 0, 1) |
|
187
|
|
|
|
|
188
|
|
|
data = pd.DataFrame(data).copy() |
|
189
|
|
|
data_cleaned = drop_missing(data, drop_threshold_cols, drop_threshold_rows) |
|
190
|
|
|
|
|
191
|
|
|
single_val_cols = data_cleaned.columns[data_cleaned.nunique(dropna=False) == 1].tolist() |
|
192
|
|
|
data_cleaned = data_cleaned.drop(columns=single_val_cols) |
|
193
|
|
|
|
|
194
|
|
|
dupl_rows = None |
|
195
|
|
|
|
|
196
|
|
|
if drop_duplicates: |
|
197
|
|
|
data_cleaned, dupl_rows = _drop_duplicates(data_cleaned) |
|
198
|
|
|
if convert_dtypes: |
|
199
|
|
|
data_cleaned = convert_datatypes(data_cleaned, category=category, cat_threshold=cat_threshold, |
|
200
|
|
|
cat_exclude=cat_exclude) |
|
201
|
|
|
|
|
202
|
|
|
_diff_report(data, data_cleaned, dupl_rows=dupl_rows, single_val_cols=single_val_cols, show=show) |
|
203
|
|
|
|
|
204
|
|
|
return data_cleaned |
|
205
|
|
|
|
|
206
|
|
|
|
|
207
|
|
|
class DataCleaner(BaseEstimator, TransformerMixin): |
|
208
|
|
|
''' |
|
209
|
|
|
Wrapper for data_cleaning(). Allows data_cleaning() to be put into a pipeline with similar \ |
|
210
|
|
|
functions (e.g. using MVColHandler() or SubsetPooler()). |
|
211
|
|
|
|
|
212
|
|
|
Parameters: |
|
213
|
|
|
---------´ |
|
214
|
|
|
drop_threshold_cols: float, default 0.9 |
|
215
|
|
|
Drop columns with NA-ratio above the specified threshold. |
|
216
|
|
|
|
|
217
|
|
|
drop_threshold_rows: float, default 0.9 |
|
218
|
|
|
Drop rows with NA-ratio above the specified threshold. |
|
219
|
|
|
|
|
220
|
|
|
drop_duplicates: bool, default True |
|
221
|
|
|
Drop duplicate rows, keeping the first occurence. This step comes after the dropping of missing values. |
|
222
|
|
|
|
|
223
|
|
|
convert_dtypes: bool, default True |
|
224
|
|
|
Convert dtypes using pd.convert_dtypes(). |
|
225
|
|
|
|
|
226
|
|
|
category: bool, default True |
|
227
|
|
|
Change dtypes of columns to "category". Set threshold using cat_threshold. Requires convert_dtypes=True |
|
228
|
|
|
|
|
229
|
|
|
cat_threshold: float, default 0.03 |
|
230
|
|
|
Ratio of unique values below which categories are inferred and column dtype is changed to categorical. |
|
231
|
|
|
|
|
232
|
|
|
cat_exclude: list, default None |
|
233
|
|
|
List of columns to exclude from categorical conversion. |
|
234
|
|
|
|
|
235
|
|
|
show: {'all', 'changes', None} default 'all' |
|
236
|
|
|
Specify verbosity of the output. |
|
237
|
|
|
* 'all': Print information about the data before and after cleaning as well as information about changes. |
|
238
|
|
|
* 'changes': Print out differences in the data before and after cleaning. |
|
239
|
|
|
* None: No information about the data and the data cleaning is printed. |
|
240
|
|
|
|
|
241
|
|
|
Returns: |
|
242
|
|
|
------- |
|
243
|
|
|
data_cleaned: Pandas DataFrame |
|
244
|
|
|
''' |
|
245
|
|
|
|
|
246
|
|
|
def __init__(self, drop_threshold_cols=0.9, drop_threshold_rows=0.9, drop_duplicates=True, convert_dtypes=True, |
|
247
|
|
|
category=True, cat_threshold=0.03, cat_exclude=None, show='changes'): |
|
248
|
|
|
self.drop_threshold_cols = drop_threshold_cols |
|
249
|
|
|
self.drop_threshold_rows = drop_threshold_rows |
|
250
|
|
|
self.drop_duplicates = drop_duplicates |
|
251
|
|
|
self.convert_dtypes = convert_dtypes |
|
252
|
|
|
self.category = category |
|
253
|
|
|
self.cat_threshold = cat_threshold |
|
254
|
|
|
self.cat_exclude = cat_exclude |
|
255
|
|
|
self.show = show |
|
256
|
|
|
|
|
257
|
|
|
def fit(self, data, target=None): |
|
258
|
|
|
return self |
|
259
|
|
|
|
|
260
|
|
|
def transform(self, data, target=None): |
|
261
|
|
|
data_cleaned = data_cleaning(data, drop_threshold_cols=self.drop_threshold_cols, |
|
262
|
|
|
drop_threshold_rows=self.drop_threshold_rows, drop_duplicates=self.drop_duplicates, |
|
263
|
|
|
convert_dtypes=self.convert_dtypes, category=self.category, cat_threshold=self. |
|
264
|
|
|
cat_threshold, cat_exclude=self.cat_exclude, show=self.show) |
|
265
|
|
|
return data_cleaned |
|
266
|
|
|
|
|
267
|
|
|
|
|
268
|
|
|
def mv_col_handling(data, target=None, mv_threshold=0.1, corr_thresh_features=0.5, corr_thresh_target=0.3, |
|
269
|
|
|
return_details=False): |
|
270
|
|
|
''' |
|
271
|
|
|
Converts columns with a high ratio of missing values into binary features and eventually drops them based on \ |
|
272
|
|
|
their correlation with other features and the target variable. This function follows a three step process: |
|
273
|
|
|
- 1) Identify features with a high ratio of missing values (above 'mv_threshold'). |
|
274
|
|
|
- 2) Identify high correlations of these features among themselves and with other features in the dataset (above \ |
|
275
|
|
|
'corr_thresh_features'). |
|
276
|
|
|
- 3) Features with high ratio of missing values and high correlation among each other are dropped unless \ |
|
277
|
|
|
they correlate reasonably well with the target variable (above 'corr_thresh_target'). |
|
278
|
|
|
|
|
279
|
|
|
Note: If no target is provided, the process exits after step two and drops columns identified up to this point. |
|
280
|
|
|
|
|
281
|
|
|
Parameters |
|
282
|
|
|
---------- |
|
283
|
|
|
data: 2D dataset that can be coerced into Pandas DataFrame. |
|
284
|
|
|
|
|
285
|
|
|
target: string, list, np.array or pd.Series, default None |
|
286
|
|
|
Specify target for correlation. I.e. label column to generate only the correlations between each feature \ |
|
287
|
|
|
and the label. |
|
288
|
|
|
|
|
289
|
|
|
mv_threshold: float, default 0.1 |
|
290
|
|
|
Value between 0 <= threshold <= 1. Features with a missing-value-ratio larger than mv_threshold are candidates \ |
|
291
|
|
|
for dropping and undergo further analysis. |
|
292
|
|
|
|
|
293
|
|
|
corr_thresh_features: float, default 0.5 |
|
294
|
|
|
Value between 0 <= threshold <= 1. Maximum correlation a previously identified features (with a high mv-ratio) \ |
|
295
|
|
|
is allowed to have with another feature. If this threshold is overstepped, the feature undergoes further \ |
|
296
|
|
|
analysis. |
|
297
|
|
|
|
|
298
|
|
|
corr_thresh_target: float, default 0.3 |
|
299
|
|
|
Value between 0 <= threshold <= 1. Minimum required correlation of a remaining feature (i.e. feature with a \ |
|
300
|
|
|
high mv-ratio and high correlation to another existing feature) with the target. If this threshold is not met \ |
|
301
|
|
|
the feature is ultimately dropped. |
|
302
|
|
|
|
|
303
|
|
|
return_details: bool, default False |
|
304
|
|
|
Provdies flexibility to return intermediary results. |
|
305
|
|
|
|
|
306
|
|
|
Returns |
|
307
|
|
|
------- |
|
308
|
|
|
data: Updated Pandas DataFrame |
|
309
|
|
|
|
|
310
|
|
|
optional: |
|
311
|
|
|
cols_mv: Columns with missing values included in the analysis |
|
312
|
|
|
drop_cols: List of dropped columns |
|
313
|
|
|
''' |
|
314
|
|
|
|
|
315
|
|
|
# Validate Inputs |
|
316
|
|
|
_validate_input_range(mv_threshold, 'mv_threshold', 0, 1) |
|
317
|
|
|
_validate_input_range(corr_thresh_features, 'corr_thresh_features', 0, 1) |
|
318
|
|
|
_validate_input_range(corr_thresh_target, 'corr_thresh_target', 0, 1) |
|
319
|
|
|
|
|
320
|
|
|
data = pd.DataFrame(data).copy() |
|
321
|
|
|
data_local = data.copy() |
|
322
|
|
|
mv_ratios = _missing_vals(data_local)['mv_cols_ratio'] |
|
323
|
|
|
cols_mv = mv_ratios[mv_ratios > mv_threshold].index.tolist() |
|
324
|
|
|
data_local[cols_mv] = data_local[cols_mv].applymap(lambda x: 1 if not pd.isnull(x) else x).fillna(0) |
|
325
|
|
|
|
|
326
|
|
|
high_corr_features = [] |
|
327
|
|
|
data_temp = data_local.copy() |
|
328
|
|
|
for col in cols_mv: |
|
329
|
|
|
corrmat = corr_mat(data_temp, colored=False) |
|
330
|
|
|
if abs(corrmat[col]).nlargest(2)[1] > corr_thresh_features: |
|
331
|
|
|
high_corr_features.append(col) |
|
332
|
|
|
data_temp = data_temp.drop(columns=[col]) |
|
333
|
|
|
|
|
334
|
|
|
drop_cols = [] |
|
335
|
|
|
if target is None: |
|
336
|
|
|
data = data.drop(columns=high_corr_features) |
|
337
|
|
|
else: |
|
338
|
|
|
corrs = corr_mat(data_local, target=target, colored=False).loc[high_corr_features] |
|
339
|
|
|
drop_cols = corrs.loc[abs(corrs.iloc[:, 0]) < corr_thresh_target].index.tolist() |
|
340
|
|
|
data = data.drop(columns=drop_cols) |
|
341
|
|
|
|
|
342
|
|
|
if return_details: |
|
343
|
|
|
return data, cols_mv, drop_cols |
|
344
|
|
|
|
|
345
|
|
|
return data |
|
346
|
|
|
|
|
347
|
|
|
|
|
348
|
|
|
class MVColHandler(BaseEstimator, TransformerMixin): |
|
349
|
|
|
''' |
|
350
|
|
|
Wrapper for mv_col_handling(). Allows mv_col_handling() to be put into a pipeline with similar \ |
|
351
|
|
|
functions (e.g. using DataCleaner() or SubsetPooler()). |
|
352
|
|
|
|
|
353
|
|
|
Parameters |
|
354
|
|
|
---------- |
|
355
|
|
|
target: string, list, np.array or pd.Series, default None |
|
356
|
|
|
Specify target for correlation. E.g. label column to generate only the correlations between each feature \ |
|
357
|
|
|
and the label. |
|
358
|
|
|
|
|
359
|
|
|
mv_threshold: float, default 0.1 |
|
360
|
|
|
Value between 0 <= threshold <= 1. Features with a missing-value-ratio larger than mv_threshold are candidates \ |
|
361
|
|
|
for dropping and undergo further analysis. |
|
362
|
|
|
|
|
363
|
|
|
corr_thresh_features: float, default 0.6 |
|
364
|
|
|
Value between 0 <= threshold <= 1. Maximum correlation a previously identified features with a high mv-ratio is\ |
|
365
|
|
|
allowed to have with another feature. If this threshold is overstepped, the feature undergoes further analysis. |
|
366
|
|
|
|
|
367
|
|
|
corr_thresh_target: float, default 0.3 |
|
368
|
|
|
Value between 0 <= threshold <= 1. Minimum required correlation of a remaining feature (i.e. feature with a \ |
|
369
|
|
|
high mv-ratio and high correlation to another existing feature) with the target. If this threshold is not met \ |
|
370
|
|
|
the feature is ultimately dropped. |
|
371
|
|
|
|
|
372
|
|
|
return_details: bool, default True |
|
373
|
|
|
Provdies flexibility to return intermediary results. |
|
374
|
|
|
|
|
375
|
|
|
Returns |
|
376
|
|
|
------- |
|
377
|
|
|
data: Updated Pandas DataFrame |
|
378
|
|
|
''' |
|
379
|
|
|
|
|
380
|
|
|
def __init__(self, target=None, mv_threshold=0.1, corr_thresh_features=0.6, corr_thresh_target=0.3, |
|
381
|
|
|
return_details=True): |
|
382
|
|
|
self.target = target |
|
383
|
|
|
self.mv_threshold = mv_threshold |
|
384
|
|
|
self.corr_thresh_features = corr_thresh_features |
|
385
|
|
|
self.corr_thresh_target = corr_thresh_target |
|
386
|
|
|
self.return_details = return_details |
|
387
|
|
|
|
|
388
|
|
|
def fit(self, data, target=None): |
|
389
|
|
|
return self |
|
390
|
|
|
|
|
391
|
|
|
def transform(self, data, target=None): |
|
392
|
|
|
data, cols_mv, dropped_cols = mv_col_handling(data, target=self.target, mv_threshold=self.mv_threshold, |
|
393
|
|
|
corr_thresh_features=self.corr_thresh_features, |
|
394
|
|
|
corr_thresh_target=self.corr_thresh_target, |
|
395
|
|
|
return_details=self.return_details) |
|
396
|
|
|
|
|
397
|
|
|
print(f'\nFeatures with MV-ratio > {self.mv_threshold}: {len(cols_mv)}') |
|
398
|
|
|
print('Features dropped:', len(dropped_cols), dropped_cols) |
|
399
|
|
|
|
|
400
|
|
|
return data |
|
401
|
|
|
|
|
402
|
|
|
|
|
403
|
|
|
def pool_duplicate_subsets(data, col_dupl_thresh=0.2, subset_thresh=0.2, min_col_pool=3, exclude=None, |
|
404
|
|
|
return_details=False): |
|
405
|
|
|
''' |
|
406
|
|
|
Checks for duplicates in subsets of columns and pools them. This can reduce the number of columns in the data \ |
|
407
|
|
|
without loosing much information. Suitable columns are combined to subsets and tested for duplicates. In case \ |
|
408
|
|
|
sufficient duplicates can be found, the respective columns are aggregated into a 'pooled_var' column. Identical \ |
|
409
|
|
|
numbers in the 'pooled_var' column indicate identical information in the respective rows. |
|
410
|
|
|
|
|
411
|
|
|
Parameters |
|
412
|
|
|
---------- |
|
413
|
|
|
data: 2D dataset that can be coerced into Pandas DataFrame. |
|
414
|
|
|
|
|
415
|
|
|
col_dupl_thresh: float, default 0.2 |
|
416
|
|
|
Columns with a ratio of duplicates higher than 'col_dupl_thresh' are considered in the further analysis. \ |
|
417
|
|
|
Columns with a lower ratio are not considered for pooling. |
|
418
|
|
|
|
|
419
|
|
|
subset_thresh: float, default 0.2 |
|
420
|
|
|
The first subset with a duplicate threshold higher than 'subset_thresh' is chosen and aggregated. If no subset \ |
|
421
|
|
|
reaches the threshold, the algorithm continues with continuously smaller subsets until 'min_col_pool' is \ |
|
422
|
|
|
reached. |
|
423
|
|
|
|
|
424
|
|
|
min_col_pool: integer, default 3 |
|
425
|
|
|
Minimum number of columns to pool. The algorithm attempts to combine as many columns as possible to suitable \ |
|
426
|
|
|
subsets and stops when 'min_col_pool' is reached. |
|
427
|
|
|
|
|
428
|
|
|
exclude. list, default None |
|
429
|
|
|
List of column names to be excluded from the analysis. These columns are passed through without modification. |
|
430
|
|
|
|
|
431
|
|
|
return_details: bool, default False |
|
432
|
|
|
Provdies flexibility to return intermediary results. |
|
433
|
|
|
|
|
434
|
|
|
Returns: |
|
435
|
|
|
------- |
|
436
|
|
|
data: pd.DataFrame |
|
437
|
|
|
|
|
438
|
|
|
optional: |
|
439
|
|
|
subset_cols: List of columns used as subset. |
|
440
|
|
|
''' |
|
441
|
|
|
|
|
442
|
|
|
# Input validation |
|
443
|
|
|
_validate_input_range(col_dupl_thresh, 'col_dupl_thresh', 0, 1) |
|
444
|
|
|
_validate_input_range(subset_thresh, 'subset_thresh', 0, 1) |
|
445
|
|
|
_validate_input_range(min_col_pool, 'min_col_pool', 0, data.shape[1]) |
|
446
|
|
|
|
|
447
|
|
|
excluded_cols = [] |
|
448
|
|
|
if exclude is not None: |
|
449
|
|
|
excluded_cols = data[exclude] |
|
450
|
|
|
data = data.drop(columns=exclude) |
|
451
|
|
|
|
|
452
|
|
|
subset_cols = [] |
|
453
|
|
|
for i in range(data.shape[1]+1-min_col_pool): |
|
454
|
|
|
check_list = [col for col in data.columns if data.duplicated(subset=col).mean() > col_dupl_thresh] |
|
455
|
|
|
|
|
456
|
|
|
if len(check_list) > 0: |
|
457
|
|
|
combinations = itertools.combinations(check_list, len(check_list)-i) |
|
458
|
|
|
else: |
|
459
|
|
|
continue |
|
460
|
|
|
|
|
461
|
|
|
ratios = [data[list(comb)].duplicated().mean() for comb in combinations] |
|
462
|
|
|
|
|
463
|
|
|
max_ratio = max(ratios) |
|
464
|
|
|
max_idx = np.argmax(ratios) |
|
465
|
|
|
|
|
466
|
|
|
if max_ratio > subset_thresh: |
|
467
|
|
|
best_subset = itertools.islice(itertools.combinations( |
|
468
|
|
|
check_list, len(check_list)-i), max_idx, max_idx+1) |
|
469
|
|
|
best_subset = data[list(list(best_subset)[0])] |
|
470
|
|
|
subset_cols = best_subset.columns.tolist() |
|
471
|
|
|
|
|
472
|
|
|
unique_subset = best_subset.drop_duplicates().reset_index().rename(columns={'index': 'pooled_vars'}) |
|
473
|
|
|
data = data.merge(unique_subset, how='left', on=best_subset.columns.tolist() |
|
474
|
|
|
).drop(columns=best_subset.columns.tolist()) |
|
475
|
|
|
data.index = pd.RangeIndex(len(data)) |
|
476
|
|
|
break |
|
477
|
|
|
|
|
478
|
|
|
data = pd.concat([data, pd.DataFrame(excluded_cols)], axis=1) |
|
479
|
|
|
|
|
480
|
|
|
if return_details: |
|
481
|
|
|
return data, subset_cols |
|
482
|
|
|
|
|
483
|
|
|
return data |
|
484
|
|
|
|
|
485
|
|
|
|
|
486
|
|
|
class SubsetPooler(BaseEstimator, TransformerMixin): |
|
487
|
|
|
''' |
|
488
|
|
|
Wrapper for pool_duplicate_subsets(). Allows pool_duplicate_subsets() to be put into a pipeline with similar \ |
|
489
|
|
|
functions (e.g. using DataCleaner() or MVColHandler()). |
|
490
|
|
|
|
|
491
|
|
|
Parameters |
|
492
|
|
|
---------- |
|
493
|
|
|
col_dupl_ratio: float, default 0.2 |
|
494
|
|
|
Columns with a ratio of duplicates higher than 'col_dupl_ratio' are considered in the further analysis. \ |
|
495
|
|
|
Columns with a lower ratio are not considered for pooling. |
|
496
|
|
|
|
|
497
|
|
|
dupl_thresh: float, default 0.2 |
|
498
|
|
|
The first subset with a duplicate threshold higher than 'dupl_thresh' is chosen and aggregated. If no subset \ |
|
499
|
|
|
reaches the threshold, the algorithm continues with continuously smaller subsets until 'min_col_pool' is \ |
|
500
|
|
|
reached. |
|
501
|
|
|
|
|
502
|
|
|
min_col_pool: integer, default 3 |
|
503
|
|
|
Minimum number of columns to pool. The algorithm attempts to combine as many columns as possible to suitable \ |
|
504
|
|
|
subsets and stops when 'min_col_pool' is reached. |
|
505
|
|
|
|
|
506
|
|
|
return_details: bool, default False |
|
507
|
|
|
Provdies flexibility to return intermediary results. |
|
508
|
|
|
|
|
509
|
|
|
Returns: |
|
510
|
|
|
------- |
|
511
|
|
|
data: pd.DataFrame |
|
512
|
|
|
''' |
|
513
|
|
|
|
|
514
|
|
|
def __init__(self, col_dupl_thresh=0.2, subset_thresh=0.2, min_col_pool=3, return_details=True): |
|
515
|
|
|
self.col_dupl_thresh = col_dupl_thresh |
|
516
|
|
|
self.subset_thresh = subset_thresh |
|
517
|
|
|
self.min_col_pool = min_col_pool |
|
518
|
|
|
self.return_details = return_details |
|
519
|
|
|
|
|
520
|
|
|
def fit(self, data, target=None): |
|
521
|
|
|
return self |
|
522
|
|
|
|
|
523
|
|
|
def transform(self, data, target=None): |
|
524
|
|
|
data, subset_cols = pool_duplicate_subsets( |
|
525
|
|
|
data, col_dupl_thresh=0.2, subset_thresh=0.2, min_col_pool=3, return_details=True) |
|
526
|
|
|
|
|
527
|
|
|
print('Combined columns:', len(subset_cols), subset_cols) |
|
528
|
|
|
|
|
529
|
|
|
return data |
|
530
|
|
|
|
akanz1 /
klib
