|
1
|
|
|
""" |
|
2
|
|
|
Functions for data cleaning. |
|
3
|
|
|
|
|
4
|
|
|
:author: Andreas Kanz |
|
5
|
|
|
""" |
|
6
|
|
|
|
|
7
|
|
|
# Imports |
|
8
|
1 |
|
import itertools |
|
9
|
1 |
|
import numpy as np |
|
10
|
1 |
|
import pandas as pd |
|
11
|
1 |
|
import re |
|
12
|
1 |
|
from sklearn.base import BaseEstimator, TransformerMixin |
|
13
|
1 |
|
from typing import List, Optional, Union |
|
14
|
|
|
|
|
15
|
1 |
|
from klib.describe import corr_mat |
|
16
|
1 |
|
from klib.utils import ( |
|
17
|
|
|
_diff_report, |
|
18
|
|
|
_drop_duplicates, |
|
19
|
|
|
_missing_vals, |
|
20
|
|
|
_validate_input_bool, |
|
21
|
|
|
_validate_input_range, |
|
22
|
|
|
) |
|
23
|
|
|
|
|
24
|
1 |
|
__all__ = [ |
|
25
|
|
|
"clean_column_names", |
|
26
|
|
|
"convert_datatypes", |
|
27
|
|
|
"data_cleaning", |
|
28
|
|
|
"drop_missing", |
|
29
|
|
|
"mv_col_handling", |
|
30
|
|
|
] |
|
31
|
|
|
|
|
32
|
|
|
|
|
33
|
1 |
|
def optimize_ints(data: Union[pd.Series, pd.DataFrame]) -> pd.DataFrame: |
|
34
|
1 |
|
data = pd.DataFrame(data).copy() |
|
35
|
1 |
|
ints = data.select_dtypes(include=["int64"]).columns.tolist() |
|
36
|
1 |
|
data[ints] = data[ints].apply(pd.to_numeric, downcast="integer") |
|
37
|
1 |
|
return data |
|
38
|
|
|
|
|
39
|
|
|
|
|
40
|
1 |
|
def optimize_floats(data: Union[pd.Series, pd.DataFrame]) -> pd.DataFrame: |
|
41
|
1 |
|
data = pd.DataFrame(data).copy() |
|
42
|
1 |
|
floats = data.select_dtypes(include=["float64"]).columns.tolist() |
|
43
|
1 |
|
data[floats] = data[floats].apply(pd.to_numeric, downcast="float") |
|
44
|
1 |
|
return data |
|
45
|
|
|
|
|
46
|
|
|
|
|
47
|
1 |
|
def clean_column_names(data: pd.DataFrame, hints: bool = True) -> pd.DataFrame: |
|
48
|
|
|
"""Clean the column names of the provided Pandas Dataframe and optionally \ |
|
49
|
|
|
provides hints on duplicate and long column names. |
|
50
|
|
|
|
|
51
|
|
|
Parameters |
|
52
|
|
|
---------- |
|
53
|
|
|
data : pd.DataFrame |
|
54
|
|
|
Original Dataframe with columns to be cleaned |
|
55
|
|
|
hints : bool, optional |
|
56
|
|
|
Print out hints on column name duplication and colum name length, by default \ |
|
57
|
|
|
True |
|
58
|
|
|
|
|
59
|
|
|
Returns |
|
60
|
|
|
------- |
|
61
|
|
|
pd.DataFrame |
|
62
|
|
|
Pandas DataFrame with cleaned column names |
|
63
|
|
|
""" |
|
64
|
1 |
|
_validate_input_bool(hints, "hints") |
|
65
|
|
|
|
|
66
|
|
|
# Handle CamelCase |
|
67
|
1 |
|
for i, col in enumerate(data.columns): |
|
68
|
1 |
|
matches = re.findall(re.compile("[a-z][A-Z]"), col) |
|
69
|
1 |
|
column = col |
|
70
|
1 |
|
for match in matches: |
|
71
|
1 |
|
column = column.replace(match, match[0] + "_" + match[1]) |
|
72
|
1 |
|
data.rename(columns={data.columns[i]: column}, inplace=True) |
|
73
|
|
|
|
|
74
|
1 |
|
data.columns = ( |
|
75
|
|
|
data.columns.str.replace("\n", "_", regex=False) |
|
76
|
|
|
.str.replace("(", "_", regex=False) |
|
77
|
|
|
.str.replace(")", "_", regex=False) |
|
78
|
|
|
.str.replace("'", "_", regex=False) |
|
79
|
|
|
.str.replace('"', "_", regex=False) |
|
80
|
|
|
.str.replace(".", "_", regex=False) |
|
81
|
|
|
.str.replace("-", "_", regex=False) |
|
82
|
|
|
.str.replace(r"[!?:;/]", "_", regex=True) |
|
83
|
|
|
.str.replace("+", "_plus_", regex=False) |
|
84
|
|
|
.str.replace("*", "_times_", regex=False) |
|
85
|
|
|
.str.replace("<", "_smaller", regex=False) |
|
86
|
|
|
.str.replace(">", "_larger_", regex=False) |
|
87
|
|
|
.str.replace("=", "_equal_", regex=False) |
|
88
|
|
|
.str.replace("ä", "ae", regex=False) |
|
89
|
|
|
.str.replace("ö", "oe", regex=False) |
|
90
|
|
|
.str.replace("ü", "ue", regex=False) |
|
91
|
|
|
.str.replace("ß", "ss", regex=False) |
|
92
|
|
|
.str.replace("%", "_percent_", regex=False) |
|
93
|
|
|
.str.replace("$", "_dollar_", regex=False) |
|
94
|
|
|
.str.replace("€", "_euro_", regex=False) |
|
95
|
|
|
.str.replace("@", "_at_", regex=False) |
|
96
|
|
|
.str.replace("#", "_hash_", regex=False) |
|
97
|
|
|
.str.replace("&", "_and_", regex=False) |
|
98
|
|
|
.str.replace(r"\s+", "_", regex=True) |
|
99
|
|
|
.str.replace(r"_+", "_", regex=True) |
|
100
|
|
|
.str.strip("_") |
|
101
|
|
|
.str.lower() |
|
102
|
|
|
) |
|
103
|
|
|
|
|
104
|
1 |
|
dupl_idx = [i for i, x in enumerate(data.columns.duplicated()) if x] |
|
105
|
1 |
|
if dupl_idx: |
|
106
|
1 |
|
dupl_before = data.columns[dupl_idx].tolist() |
|
107
|
1 |
|
data.columns = [ |
|
108
|
|
|
col if col not in data.columns[:i] else col + "_" + str(i) |
|
109
|
|
|
for i, col in enumerate(data.columns) |
|
110
|
|
|
] |
|
111
|
1 |
|
if hints: |
|
112
|
1 |
|
print( |
|
113
|
|
|
f"Duplicate column names detected! Columns with index {dupl_idx} and " |
|
114
|
|
|
f"names {dupl_before}) have been renamed to " |
|
115
|
|
|
f"{data.columns[dupl_idx].tolist()}." |
|
116
|
|
|
) |
|
117
|
|
|
|
|
118
|
1 |
|
long_col_names = [x for x in data.columns if len(x) > 25] |
|
119
|
1 |
|
if long_col_names and hints: |
|
120
|
1 |
|
print( |
|
121
|
|
|
"Long column names detected (>25 characters). Consider renaming the " |
|
122
|
|
|
f"following columns {long_col_names}." |
|
123
|
|
|
) |
|
124
|
|
|
|
|
125
|
1 |
|
return data |
|
126
|
|
|
|
|
127
|
|
|
|
|
128
|
1 |
|
def convert_datatypes( |
|
129
|
|
|
data: pd.DataFrame, |
|
130
|
|
|
category: bool = True, |
|
131
|
|
|
cat_threshold: float = 0.05, |
|
132
|
|
|
cat_exclude: Optional[List[Union[str, int]]] = None, |
|
133
|
|
|
) -> pd.DataFrame: |
|
134
|
|
|
"""Convert columns to best possible dtypes using dtypes supporting pd.NA. |
|
135
|
|
|
|
|
136
|
|
|
Temporarily not converting to integers due to an issue in pandas. This is expected \ |
|
137
|
|
|
to be fixed in pandas 1.1. See https://github.com/pandas-dev/pandas/issues/33803 |
|
138
|
|
|
|
|
139
|
|
|
Parameters |
|
140
|
|
|
---------- |
|
141
|
|
|
data : pd.DataFrame |
|
142
|
|
|
2D dataset that can be coerced into Pandas DataFrame |
|
143
|
|
|
category : bool, optional |
|
144
|
|
|
Change dtypes of columns with dtype "object" to "category". Set threshold \ |
|
145
|
|
|
using cat_threshold or exclude columns using cat_exclude, by default True |
|
146
|
|
|
cat_threshold : float, optional |
|
147
|
|
|
Ratio of unique values below which categories are inferred and column dtype is \ |
|
148
|
|
|
changed to categorical, by default 0.05 |
|
149
|
|
|
cat_exclude : Optional[List[Union[str, int]]], optional |
|
150
|
|
|
List of columns to exclude from categorical conversion, by default None |
|
151
|
|
|
|
|
152
|
|
|
Returns |
|
153
|
|
|
------- |
|
154
|
|
|
pd.DataFrame |
|
155
|
|
|
Pandas DataFrame with converted Datatypes |
|
156
|
|
|
""" |
|
157
|
|
|
# Validate Inputs |
|
158
|
1 |
|
_validate_input_bool(category, "Category") |
|
159
|
1 |
|
_validate_input_range(cat_threshold, "cat_threshold", 0, 1) |
|
160
|
|
|
|
|
161
|
1 |
|
cat_exclude = [] if cat_exclude is None else cat_exclude.copy() |
|
162
|
|
|
|
|
163
|
1 |
|
data = pd.DataFrame(data).copy() |
|
164
|
1 |
|
for col in data.columns: |
|
165
|
1 |
|
unique_vals_ratio = data[col].nunique(dropna=False) / data.shape[0] |
|
166
|
1 |
|
if ( |
|
167
|
|
|
category |
|
168
|
|
|
and unique_vals_ratio < cat_threshold |
|
169
|
|
|
and col not in cat_exclude |
|
170
|
|
|
and data[col].dtype == "object" |
|
171
|
|
|
): |
|
172
|
1 |
|
data[col] = data[col].astype("category") |
|
173
|
|
|
|
|
174
|
1 |
|
data[col] = data[col].convert_dtypes( |
|
175
|
|
|
infer_objects=True, |
|
176
|
|
|
convert_string=True, |
|
177
|
|
|
convert_integer=False, |
|
178
|
|
|
convert_boolean=True, |
|
179
|
|
|
) |
|
180
|
|
|
|
|
181
|
1 |
|
data = optimize_ints(data) |
|
182
|
1 |
|
data = optimize_floats(data) |
|
183
|
|
|
|
|
184
|
1 |
|
return data |
|
185
|
|
|
|
|
186
|
|
|
|
|
187
|
1 |
|
def drop_missing( |
|
188
|
|
|
data: pd.DataFrame, |
|
189
|
|
|
drop_threshold_cols: float = 1, |
|
190
|
|
|
drop_threshold_rows: float = 1, |
|
191
|
|
|
col_exclude: Optional[List[str]] = None, |
|
192
|
|
|
) -> pd.DataFrame: |
|
193
|
|
|
"""Drop completely empty columns and rows by default and optionally provides \ |
|
194
|
|
|
flexibility to loosen restrictions to drop additional non-empty columns and \ |
|
195
|
|
|
rows based on the fraction of NA-values. |
|
196
|
|
|
|
|
197
|
|
|
Parameters |
|
198
|
|
|
---------- |
|
199
|
|
|
data : pd.DataFrame |
|
200
|
|
|
2D dataset that can be coerced into Pandas DataFrame |
|
201
|
|
|
drop_threshold_cols : float, optional |
|
202
|
|
|
Drop columns with NA-ratio equal to or above the specified threshold, by \ |
|
203
|
|
|
default 1 |
|
204
|
|
|
drop_threshold_rows : float, optional |
|
205
|
|
|
Drop rows with NA-ratio equal to or above the specified threshold, by default 1 |
|
206
|
|
|
col_exclude : Optional[List[str]], optional |
|
207
|
|
|
Specify a list of columns to exclude from dropping. The excluded columns do \ |
|
208
|
|
|
not affect the drop thresholds, by default None |
|
209
|
|
|
|
|
210
|
|
|
Returns |
|
211
|
|
|
------- |
|
212
|
|
|
pd.DataFrame |
|
213
|
|
|
Pandas DataFrame without any empty columns or rows |
|
214
|
|
|
|
|
215
|
|
|
Notes |
|
216
|
|
|
----- |
|
217
|
|
|
Columns are dropped first |
|
218
|
|
|
""" |
|
219
|
|
|
# Validate Inputs |
|
220
|
1 |
|
_validate_input_range(drop_threshold_cols, "drop_threshold_cols", 0, 1) |
|
221
|
1 |
|
_validate_input_range(drop_threshold_rows, "drop_threshold_rows", 0, 1) |
|
222
|
|
|
|
|
223
|
1 |
|
col_exclude = [] if col_exclude is None else col_exclude.copy() |
|
224
|
1 |
|
data_exclude = data[col_exclude] |
|
225
|
|
|
|
|
226
|
1 |
|
data = pd.DataFrame(data).copy() |
|
227
|
|
|
|
|
228
|
1 |
|
data_dropped = data.drop(columns=col_exclude, errors="ignore") |
|
229
|
1 |
|
data_dropped = data_dropped.drop( |
|
230
|
|
|
columns=data_dropped.loc[ |
|
231
|
|
|
:, _missing_vals(data)["mv_cols_ratio"] > drop_threshold_cols |
|
232
|
|
|
].columns |
|
233
|
|
|
).dropna(axis=1, how="all") |
|
234
|
|
|
|
|
235
|
1 |
|
data = pd.concat([data_dropped, data_exclude], axis=1) |
|
236
|
|
|
|
|
237
|
1 |
|
return data.drop( |
|
238
|
|
|
index=data.loc[ |
|
239
|
|
|
_missing_vals(data)["mv_rows_ratio"] > drop_threshold_rows, : |
|
240
|
|
|
].index |
|
241
|
|
|
).dropna(axis=0, how="all") |
|
242
|
|
|
|
|
243
|
|
|
|
|
244
|
1 |
|
def data_cleaning( |
|
245
|
|
|
data: pd.DataFrame, |
|
246
|
|
|
drop_threshold_cols: float = 0.9, |
|
247
|
|
|
drop_threshold_rows: float = 0.9, |
|
248
|
|
|
drop_duplicates: bool = True, |
|
249
|
|
|
convert_dtypes: bool = True, |
|
250
|
|
|
col_exclude: Optional[List[str]] = None, |
|
251
|
|
|
category: bool = True, |
|
252
|
|
|
cat_threshold: float = 0.03, |
|
253
|
|
|
cat_exclude: Optional[List[Union[str, int]]] = None, |
|
254
|
|
|
clean_col_names: bool = True, |
|
255
|
|
|
show: str = "changes", |
|
256
|
|
|
) -> pd.DataFrame: |
|
257
|
|
|
"""Perform initial data cleaning tasks on a dataset, such as dropping single \ |
|
258
|
|
|
valued and empty rows, empty columns as well as optimizing the datatypes. |
|
259
|
|
|
|
|
260
|
|
|
Parameters |
|
261
|
|
|
---------- |
|
262
|
|
|
data : pd.DataFrame |
|
263
|
|
|
2D dataset that can be coerced into Pandas DataFrame |
|
264
|
|
|
drop_threshold_cols : float, optional |
|
265
|
|
|
Drop columns with NA-ratio equal to or above the specified threshold, by \ |
|
266
|
|
|
default 0.9 |
|
267
|
|
|
drop_threshold_rows : float, optional |
|
268
|
|
|
Drop rows with NA-ratio equal to or above the specified threshold, by \ |
|
269
|
|
|
default 0.9 |
|
270
|
|
|
drop_duplicates : bool, optional |
|
271
|
|
|
Drop duplicate rows, keeping the first occurence. This step comes after the \ |
|
272
|
|
|
dropping of missing values, by default True |
|
273
|
|
|
convert_dtypes : bool, optional |
|
274
|
|
|
Convert dtypes using pd.convert_dtypes(), by default True |
|
275
|
|
|
col_exclude : Optional[List[str]], optional |
|
276
|
|
|
Specify a list of columns to exclude from dropping, by default None |
|
277
|
|
|
category : bool, optional |
|
278
|
|
|
Enable changing dtypes of "object" columns to "category". Set threshold using \ |
|
279
|
|
|
cat_threshold. Requires convert_dtypes=True, by default True |
|
280
|
|
|
cat_threshold : float, optional |
|
281
|
|
|
Ratio of unique values below which categories are inferred and column dtype is \ |
|
282
|
|
|
changed to categorical, by default 0.03 |
|
283
|
|
|
cat_exclude : Optional[List[str]], optional |
|
284
|
|
|
List of columns to exclude from categorical conversion, by default None |
|
285
|
|
|
clean_column_names: bool, optional |
|
286
|
|
|
Cleans the column names and provides hints on duplicate and long names, by \ |
|
287
|
|
|
default True |
|
288
|
|
|
show : str, optional |
|
289
|
|
|
{"all", "changes", None}, by default "changes" |
|
290
|
|
|
Specify verbosity of the output: |
|
291
|
|
|
|
|
292
|
|
|
* "all": Print information about the data before and after cleaning as \ |
|
293
|
|
|
well as information about changes and memory usage (deep). Please be \ |
|
294
|
|
|
aware, that this can slow down the function by quite a bit. |
|
295
|
|
|
* "changes": Print out differences in the data before and after cleaning. |
|
296
|
|
|
* None: No information about the data and the data cleaning is printed. |
|
297
|
|
|
|
|
298
|
|
|
Returns |
|
299
|
|
|
------- |
|
300
|
|
|
pd.DataFrame |
|
301
|
|
|
Cleaned Pandas DataFrame |
|
302
|
|
|
|
|
303
|
|
|
See Also |
|
304
|
|
|
-------- |
|
305
|
|
|
convert_datatypes: Convert columns to best possible dtypes. |
|
306
|
|
|
drop_missing : Flexibly drop columns and rows. |
|
307
|
|
|
_memory_usage: Gives the total memory usage in megabytes. |
|
308
|
|
|
_missing_vals: Metrics about missing values in the dataset. |
|
309
|
|
|
|
|
310
|
|
|
Notes |
|
311
|
|
|
----- |
|
312
|
|
|
The category dtype is not grouped in the summary, unless it contains exactly the \ |
|
313
|
|
|
same categories. |
|
314
|
|
|
""" |
|
315
|
|
|
# Validate Inputs |
|
316
|
1 |
|
_validate_input_range(drop_threshold_cols, "drop_threshold_cols", 0, 1) |
|
317
|
1 |
|
_validate_input_range(drop_threshold_rows, "drop_threshold_rows", 0, 1) |
|
318
|
1 |
|
_validate_input_bool(drop_duplicates, "drop_duplicates") |
|
319
|
1 |
|
_validate_input_bool(convert_dtypes, "convert_datatypes") |
|
320
|
1 |
|
_validate_input_bool(category, "category") |
|
321
|
1 |
|
_validate_input_range(cat_threshold, "cat_threshold", 0, 1) |
|
322
|
|
|
|
|
323
|
1 |
|
data = pd.DataFrame(data).copy() |
|
324
|
1 |
|
data_cleaned = drop_missing( |
|
325
|
|
|
data, drop_threshold_cols, drop_threshold_rows, col_exclude=col_exclude |
|
326
|
|
|
) |
|
327
|
|
|
|
|
328
|
1 |
|
if clean_col_names: |
|
329
|
1 |
|
data_cleaned = clean_column_names(data_cleaned) |
|
330
|
|
|
|
|
331
|
1 |
|
single_val_cols = data_cleaned.columns[ |
|
332
|
|
|
data_cleaned.nunique(dropna=False) == 1 |
|
333
|
|
|
].tolist() |
|
334
|
1 |
|
data_cleaned = data_cleaned.drop(columns=single_val_cols) |
|
335
|
|
|
|
|
336
|
1 |
|
dupl_rows = None |
|
337
|
|
|
|
|
338
|
1 |
|
if drop_duplicates: |
|
339
|
1 |
|
data_cleaned, dupl_rows = _drop_duplicates(data_cleaned) |
|
340
|
1 |
|
if convert_dtypes: |
|
341
|
1 |
|
data_cleaned = convert_datatypes( |
|
342
|
|
|
data_cleaned, |
|
343
|
|
|
category=category, |
|
344
|
|
|
cat_threshold=cat_threshold, |
|
345
|
|
|
cat_exclude=cat_exclude, |
|
346
|
|
|
) |
|
347
|
|
|
|
|
348
|
1 |
|
_diff_report( |
|
349
|
|
|
data, |
|
350
|
|
|
data_cleaned, |
|
351
|
|
|
dupl_rows=dupl_rows, |
|
352
|
|
|
single_val_cols=single_val_cols, |
|
353
|
|
|
show=show, |
|
354
|
|
|
) |
|
355
|
|
|
|
|
356
|
1 |
|
return data_cleaned |
|
357
|
|
|
|
|
358
|
|
|
|
|
359
|
1 |
|
class DataCleaner(BaseEstimator, TransformerMixin): |
|
360
|
|
|
"""Wrapper for data_cleaning(). Allows data_cleaning() to be put into a pipeline \ |
|
361
|
|
|
with similar functions (e.g. using MVColHandler() or SubsetPooler()). |
|
362
|
|
|
|
|
363
|
|
|
Parameters |
|
364
|
|
|
---------- |
|
365
|
|
|
drop_threshold_cols: float, default 0.9 |
|
366
|
|
|
Drop columns with NA-ratio equal to or above the specified threshold. |
|
367
|
|
|
drop_threshold_rows: float, default 0.9 |
|
368
|
|
|
Drop rows with NA-ratio equal to or above the specified threshold. |
|
369
|
|
|
drop_duplicates: bool, default True |
|
370
|
|
|
Drop duplicate rows, keeping the first occurence. This step comes after the \ |
|
371
|
|
|
dropping of missing values. |
|
372
|
|
|
convert_dtypes: bool, default True |
|
373
|
|
|
Convert dtypes using pd.convert_dtypes(). |
|
374
|
|
|
col_exclude: list, default None |
|
375
|
|
|
Specify a list of columns to exclude from dropping. |
|
376
|
|
|
category: bool, default True |
|
377
|
|
|
Change dtypes of columns to "category". Set threshold using cat_threshold. \ |
|
378
|
|
|
Requires convert_dtypes=True |
|
379
|
|
|
cat_threshold: float, default 0.03 |
|
380
|
|
|
Ratio of unique values below which categories are inferred and column dtype is \ |
|
381
|
|
|
changed to categorical. |
|
382
|
|
|
cat_exclude: list, default None |
|
383
|
|
|
List of columns to exclude from categorical conversion. |
|
384
|
|
|
clean_column_names: bool, optional |
|
385
|
|
|
Cleans the column names and provides hints on duplicate and long names, by \ |
|
386
|
|
|
default True |
|
387
|
|
|
show: str, optional |
|
388
|
|
|
{"all", "changes", None}, by default "changes" |
|
389
|
|
|
Specify verbosity of the output: |
|
390
|
|
|
* "all": Print information about the data before and after cleaning as \ |
|
391
|
|
|
well as information about changes and memory usage (deep). Please be \ |
|
392
|
|
|
aware, that this can slow down the function by quite a bit. |
|
393
|
|
|
* "changes": Print out differences in the data before and after cleaning. |
|
394
|
|
|
* None: No information about the data and the data cleaning is printed. |
|
395
|
|
|
|
|
396
|
|
|
Returns |
|
397
|
|
|
------- |
|
398
|
|
|
data_cleaned: Pandas DataFrame |
|
399
|
|
|
""" |
|
400
|
|
|
|
|
401
|
1 |
|
def __init__( |
|
402
|
|
|
self, |
|
403
|
|
|
drop_threshold_cols: float = 0.9, |
|
404
|
|
|
drop_threshold_rows: float = 0.9, |
|
405
|
|
|
drop_duplicates: bool = True, |
|
406
|
|
|
convert_dtypes: bool = True, |
|
407
|
|
|
col_exclude: Optional[List[str]] = None, |
|
408
|
|
|
category: bool = True, |
|
409
|
|
|
cat_threshold: float = 0.03, |
|
410
|
|
|
cat_exclude: Optional[List[Union[str, int]]] = None, |
|
411
|
|
|
clean_col_names: bool = True, |
|
412
|
|
|
show: str = "changes", |
|
413
|
|
|
): |
|
414
|
|
|
self.drop_threshold_cols = drop_threshold_cols |
|
415
|
|
|
self.drop_threshold_rows = drop_threshold_rows |
|
416
|
|
|
self.drop_duplicates = drop_duplicates |
|
417
|
|
|
self.convert_dtypes = convert_dtypes |
|
418
|
|
|
self.col_exclude = col_exclude |
|
419
|
|
|
self.category = category |
|
420
|
|
|
self.cat_threshold = cat_threshold |
|
421
|
|
|
self.cat_exclude = cat_exclude |
|
422
|
|
|
self.clean_col_names = clean_col_names |
|
423
|
|
|
self.show = show |
|
424
|
|
|
|
|
425
|
1 |
|
def fit(self, data, target=None): |
|
426
|
|
|
return self |
|
427
|
|
|
|
|
428
|
1 |
|
def transform(self, data, target=None): |
|
429
|
|
|
return data_cleaning( |
|
430
|
|
|
data, |
|
431
|
|
|
drop_threshold_cols=self.drop_threshold_cols, |
|
432
|
|
|
drop_threshold_rows=self.drop_threshold_rows, |
|
433
|
|
|
drop_duplicates=self.drop_duplicates, |
|
434
|
|
|
convert_dtypes=self.convert_dtypes, |
|
435
|
|
|
col_exclude=self.col_exclude, |
|
436
|
|
|
category=self.category, |
|
437
|
|
|
cat_threshold=self.cat_threshold, |
|
438
|
|
|
cat_exclude=self.cat_exclude, |
|
439
|
|
|
clean_col_names=self.clean_col_names, |
|
440
|
|
|
show=self.show, |
|
441
|
|
|
) |
|
442
|
|
|
|
|
443
|
|
|
|
|
444
|
1 |
|
def mv_col_handling( |
|
445
|
|
|
data: pd.DataFrame, |
|
446
|
|
|
target: Optional[Union[str, pd.Series, List]] = None, |
|
447
|
|
|
mv_threshold: float = 0.1, |
|
448
|
|
|
corr_thresh_features: float = 0.5, |
|
449
|
|
|
corr_thresh_target: float = 0.3, |
|
450
|
|
|
return_details: bool = False, |
|
451
|
|
|
) -> pd.DataFrame: |
|
452
|
|
|
"""Convert columns with a high ratio of missing values into binary features and \ |
|
453
|
|
|
eventually drops them based on their correlation with other features and the \ |
|
454
|
|
|
target variable. |
|
455
|
|
|
|
|
456
|
|
|
This function follows a three step process: |
|
457
|
|
|
- 1) Identify features with a high ratio of missing values (above 'mv_threshold'). |
|
458
|
|
|
- 2) Identify high correlations of these features among themselves and with \ |
|
459
|
|
|
other features in the dataset (above 'corr_thresh_features'). |
|
460
|
|
|
- 3) Features with high ratio of missing values and high correlation among each \ |
|
461
|
|
|
other are dropped unless they correlate reasonably well with the target \ |
|
462
|
|
|
variable (above 'corr_thresh_target'). |
|
463
|
|
|
|
|
464
|
|
|
Note: If no target is provided, the process exits after step two and drops columns \ |
|
465
|
|
|
identified up to this point. |
|
466
|
|
|
|
|
467
|
|
|
Parameters |
|
468
|
|
|
---------- |
|
469
|
|
|
data : pd.DataFrame |
|
470
|
|
|
2D dataset that can be coerced into Pandas DataFrame |
|
471
|
|
|
target : Optional[Union[str, pd.Series, List]], optional |
|
472
|
|
|
Specify target for correlation. I.e. label column to generate only the \ |
|
473
|
|
|
correlations between each feature and the label, by default None |
|
474
|
|
|
mv_threshold : float, optional |
|
475
|
|
|
Value between 0 <= threshold <= 1. Features with a missing-value-ratio larger \ |
|
476
|
|
|
than mv_threshold are candidates for dropping and undergo further analysis, by \ |
|
477
|
|
|
default 0.1 |
|
478
|
|
|
corr_thresh_features : float, optional |
|
479
|
|
|
Value between 0 <= threshold <= 1. Maximum correlation a previously identified \ |
|
480
|
|
|
features (with a high mv-ratio) is allowed to have with another feature. If \ |
|
481
|
|
|
this threshold is overstepped, the feature undergoes further analysis, by \ |
|
482
|
|
|
default 0.5 |
|
483
|
|
|
corr_thresh_target : float, optional |
|
484
|
|
|
Value between 0 <= threshold <= 1. Minimum required correlation of a remaining \ |
|
485
|
|
|
feature (i.e. feature with a high mv-ratio and high correlation to another \ |
|
486
|
|
|
existing feature) with the target. If this threshold is not met the feature is \ |
|
487
|
|
|
ultimately dropped, by default 0.3 |
|
488
|
|
|
return_details : bool, optional |
|
489
|
|
|
Provdies flexibility to return intermediary results, by default False |
|
490
|
|
|
|
|
491
|
|
|
Returns |
|
492
|
|
|
------- |
|
493
|
|
|
pd.DataFrame |
|
494
|
|
|
Updated Pandas DataFrame |
|
495
|
|
|
|
|
496
|
|
|
optional: |
|
497
|
|
|
cols_mv: Columns with missing values included in the analysis |
|
498
|
|
|
drop_cols: List of dropped columns |
|
499
|
|
|
""" |
|
500
|
|
|
# Validate Inputs |
|
501
|
|
|
_validate_input_range(mv_threshold, "mv_threshold", 0, 1) |
|
502
|
|
|
_validate_input_range(corr_thresh_features, "corr_thresh_features", 0, 1) |
|
503
|
|
|
_validate_input_range(corr_thresh_target, "corr_thresh_target", 0, 1) |
|
504
|
|
|
|
|
505
|
|
|
data = pd.DataFrame(data).copy() |
|
506
|
|
|
data_local = data.copy() |
|
507
|
|
|
mv_ratios = _missing_vals(data_local)["mv_cols_ratio"] |
|
508
|
|
|
cols_mv = mv_ratios[mv_ratios > mv_threshold].index.tolist() |
|
509
|
|
|
data_local[cols_mv] = ( |
|
510
|
|
|
data_local[cols_mv].applymap(lambda x: 1 if not pd.isnull(x) else x).fillna(0) |
|
511
|
|
|
) |
|
512
|
|
|
|
|
513
|
|
|
high_corr_features = [] |
|
514
|
|
|
data_temp = data_local.copy() |
|
515
|
|
|
for col in cols_mv: |
|
516
|
|
|
corrmat = corr_mat(data_temp, colored=False) |
|
517
|
|
|
if abs(corrmat[col]).nlargest(2)[1] > corr_thresh_features: |
|
518
|
|
|
high_corr_features.append(col) |
|
519
|
|
|
data_temp = data_temp.drop(columns=[col]) |
|
520
|
|
|
|
|
521
|
|
|
drop_cols = [] |
|
522
|
|
|
if target is None: |
|
523
|
|
|
data = data.drop(columns=high_corr_features) |
|
524
|
|
|
else: |
|
525
|
|
|
corrs = corr_mat(data_local, target=target, colored=False).loc[ |
|
526
|
|
|
high_corr_features |
|
527
|
|
|
] |
|
528
|
|
|
drop_cols = corrs.loc[abs(corrs.iloc[:, 0]) < corr_thresh_target].index.tolist() |
|
529
|
|
|
data = data.drop(columns=drop_cols) |
|
530
|
|
|
|
|
531
|
|
|
if return_details: |
|
532
|
|
|
return data, cols_mv, drop_cols |
|
533
|
|
|
|
|
534
|
|
|
return data |
|
535
|
|
|
|
|
536
|
|
|
|
|
537
|
1 |
|
class MVColHandler(BaseEstimator, TransformerMixin): |
|
538
|
|
|
"""Wrapper for mv_col_handling(). Allows mv_col_handling() to be put into a \ |
|
539
|
|
|
pipeline with similar functions (e.g. using DataCleaner() or SubsetPooler()). |
|
540
|
|
|
|
|
541
|
|
|
Parameters |
|
542
|
|
|
---------- |
|
543
|
|
|
target: string, list, np.array or pd.Series, default None |
|
544
|
|
|
Specify target for correlation. E.g. label column to generate only the \ |
|
545
|
|
|
correlations between each feature and the label. |
|
546
|
|
|
mv_threshold: float, default 0.1 |
|
547
|
|
|
Value between 0 <= threshold <= 1. Features with a missing-value-ratio larger \ |
|
548
|
|
|
than mv_threshold are candidates for dropping and undergo further analysis. |
|
549
|
|
|
corr_thresh_features: float, default 0.6 |
|
550
|
|
|
Value between 0 <= threshold <= 1. Maximum correlation a previously identified \ |
|
551
|
|
|
features with a high mv-ratio is allowed to have with another feature. If this \ |
|
552
|
|
|
threshold is overstepped, the feature undergoes further analysis. |
|
553
|
|
|
corr_thresh_target: float, default 0.3 |
|
554
|
|
|
Value between 0 <= threshold <= 1. Minimum required correlation of a remaining \ |
|
555
|
|
|
feature (i.e. feature with a high mv-ratio and high correlation to another \ |
|
556
|
|
|
existing feature) with the target. If this threshold is not met the feature is \ |
|
557
|
|
|
ultimately dropped. |
|
558
|
|
|
return_details: bool, default True |
|
559
|
|
|
Provdies flexibility to return intermediary results. |
|
560
|
|
|
|
|
561
|
|
|
Returns |
|
562
|
|
|
------- |
|
563
|
|
|
data: Updated Pandas DataFrame |
|
564
|
|
|
""" |
|
565
|
|
|
|
|
566
|
1 |
|
def __init__( |
|
567
|
|
|
self, |
|
568
|
|
|
target: Optional[Union[str, pd.Series, List]] = None, |
|
569
|
|
|
mv_threshold: float = 0.1, |
|
570
|
|
|
corr_thresh_features: float = 0.6, |
|
571
|
|
|
corr_thresh_target: float = 0.3, |
|
572
|
|
|
return_details: bool = True, |
|
573
|
|
|
): |
|
574
|
|
|
self.target = target |
|
575
|
|
|
self.mv_threshold = mv_threshold |
|
576
|
|
|
self.corr_thresh_features = corr_thresh_features |
|
577
|
|
|
self.corr_thresh_target = corr_thresh_target |
|
578
|
|
|
self.return_details = return_details |
|
579
|
|
|
|
|
580
|
1 |
|
def fit(self, data, target=None): |
|
581
|
|
|
return self |
|
582
|
|
|
|
|
583
|
1 |
|
def transform(self, data, target=None): |
|
584
|
|
|
data, cols_mv, dropped_cols = mv_col_handling( |
|
585
|
|
|
data, |
|
586
|
|
|
target=self.target, |
|
587
|
|
|
mv_threshold=self.mv_threshold, |
|
588
|
|
|
corr_thresh_features=self.corr_thresh_features, |
|
589
|
|
|
corr_thresh_target=self.corr_thresh_target, |
|
590
|
|
|
return_details=self.return_details, |
|
591
|
|
|
) |
|
592
|
|
|
|
|
593
|
|
|
print(f"\nFeatures with MV-ratio > {self.mv_threshold}: {len(cols_mv)}") |
|
594
|
|
|
print("Features dropped:", len(dropped_cols), dropped_cols) |
|
595
|
|
|
|
|
596
|
|
|
return data |
|
597
|
|
|
|
|
598
|
|
|
|
|
599
|
1 |
|
def pool_duplicate_subsets( |
|
600
|
|
|
data: pd.DataFrame, |
|
601
|
|
|
col_dupl_thresh: float = 0.2, |
|
602
|
|
|
subset_thresh: float = 0.2, |
|
603
|
|
|
min_col_pool: int = 3, |
|
604
|
|
|
exclude: Optional[List[str]] = None, |
|
605
|
|
|
return_details=False, |
|
606
|
|
|
) -> pd.DataFrame: |
|
607
|
|
|
"""Check for duplicates in subsets of columns and pools them. This can reduce \ |
|
608
|
|
|
the number of columns in the data without loosing much information. Suitable \ |
|
609
|
|
|
columns are combined to subsets and tested for duplicates. In case sufficient \ |
|
610
|
|
|
duplicates can be found, the respective columns are aggregated into a \ |
|
611
|
|
|
"pooled_var" column. Identical numbers in the "pooled_var" column indicate \ |
|
612
|
|
|
identical information in the respective rows. |
|
613
|
|
|
|
|
614
|
|
|
Note: It is advised to exclude features that provide sufficient informational \ |
|
615
|
|
|
content by themselves as well as the target column by using the "exclude" \ |
|
616
|
|
|
setting. |
|
617
|
|
|
|
|
618
|
|
|
Parameters |
|
619
|
|
|
---------- |
|
620
|
|
|
data : pd.DataFrame |
|
621
|
|
|
2D dataset that can be coerced into Pandas DataFrame |
|
622
|
|
|
col_dupl_thresh : float, optional |
|
623
|
|
|
Columns with a ratio of duplicates higher than "col_dupl_thresh" are \ |
|
624
|
|
|
considered in the further analysis. Columns with a lower ratio are not \ |
|
625
|
|
|
considered for pooling, by default 0.2 |
|
626
|
|
|
subset_thresh : float, optional |
|
627
|
|
|
The first subset with a duplicate threshold higher than "subset_thresh" is \ |
|
628
|
|
|
chosen and aggregated. If no subset reaches the threshold, the algorithm \ |
|
629
|
|
|
continues with continuously smaller subsets until "min_col_pool" is reached, \ |
|
630
|
|
|
by default 0.2 |
|
631
|
|
|
min_col_pool : int, optional |
|
632
|
|
|
Minimum number of columns to pool. The algorithm attempts to combine as many \ |
|
633
|
|
|
columns as possible to suitable subsets and stops when "min_col_pool" is \ |
|
634
|
|
|
reached, by default 3 |
|
635
|
|
|
exclude : Optional[List[str]], optional |
|
636
|
|
|
List of column names to be excluded from the analysis. These columns are \ |
|
637
|
|
|
passed through without modification, by default None |
|
638
|
|
|
return_details : bool, optional |
|
639
|
|
|
Provdies flexibility to return intermediary results, by default False |
|
640
|
|
|
|
|
641
|
|
|
Returns |
|
642
|
|
|
------- |
|
643
|
|
|
pd.DataFrame |
|
644
|
|
|
DataFrame with low cardinality columns pooled |
|
645
|
|
|
|
|
646
|
|
|
optional: |
|
647
|
|
|
subset_cols: List of columns used as subset |
|
648
|
|
|
""" |
|
649
|
|
|
# Input validation |
|
650
|
1 |
|
_validate_input_range(col_dupl_thresh, "col_dupl_thresh", 0, 1) |
|
651
|
1 |
|
_validate_input_range(subset_thresh, "subset_thresh", 0, 1) |
|
652
|
1 |
|
_validate_input_range(min_col_pool, "min_col_pool", 0, data.shape[1]) |
|
653
|
|
|
|
|
654
|
1 |
|
excluded_cols = [] |
|
655
|
1 |
|
if exclude is not None: |
|
656
|
|
|
excluded_cols = data[exclude] |
|
657
|
|
|
data = data.drop(columns=exclude) |
|
658
|
|
|
|
|
659
|
1 |
|
subset_cols = [] |
|
660
|
1 |
|
for i in range(data.shape[1] + 1 - min_col_pool): |
|
661
|
|
|
# Consider only columns with lots of duplicates |
|
662
|
1 |
|
check_list = [ |
|
663
|
|
|
col |
|
664
|
|
|
for col in data.columns |
|
665
|
|
|
if data.duplicated(subset=col).mean() > col_dupl_thresh |
|
666
|
|
|
] |
|
667
|
|
|
|
|
668
|
|
|
# Identify all possible combinations for the current interation |
|
669
|
1 |
|
if check_list: |
|
670
|
1 |
|
combinations = itertools.combinations(check_list, len(check_list) - i) |
|
671
|
|
|
else: |
|
672
|
|
|
continue |
|
673
|
|
|
|
|
674
|
|
|
# Check subsets for all possible combinations |
|
675
|
1 |
|
ratios = [ |
|
676
|
|
|
*map(lambda comb: data.duplicated(subset=list(comb)).mean(), combinations) |
|
677
|
|
|
] |
|
678
|
1 |
|
max_idx = np.argmax(ratios) |
|
679
|
|
|
|
|
680
|
1 |
|
if max(ratios) > subset_thresh: |
|
681
|
|
|
# Get the best possible iterator and process the data |
|
682
|
1 |
|
best_subset = itertools.islice( |
|
683
|
|
|
itertools.combinations(check_list, len(check_list) - i), |
|
684
|
|
|
max_idx, |
|
685
|
|
|
max_idx + 1, |
|
686
|
|
|
) |
|
687
|
|
|
|
|
688
|
1 |
|
best_subset = data[list(list(best_subset)[0])] |
|
689
|
1 |
|
subset_cols = best_subset.columns.tolist() |
|
690
|
|
|
|
|
691
|
1 |
|
unique_subset = ( |
|
692
|
|
|
best_subset.drop_duplicates() |
|
693
|
|
|
.reset_index() |
|
694
|
|
|
.rename(columns={"index": "pooled_vars"}) |
|
695
|
|
|
) |
|
696
|
1 |
|
data = data.merge(unique_subset, how="left", on=subset_cols).drop( |
|
697
|
|
|
columns=subset_cols |
|
698
|
|
|
) |
|
699
|
1 |
|
data.index = pd.RangeIndex(len(data)) |
|
700
|
1 |
|
break |
|
701
|
|
|
|
|
702
|
1 |
|
data = pd.concat([data, pd.DataFrame(excluded_cols)], axis=1) |
|
703
|
|
|
|
|
704
|
1 |
|
if return_details: |
|
705
|
|
|
return data, subset_cols |
|
706
|
|
|
|
|
707
|
1 |
|
return data |
|
708
|
|
|
|
|
709
|
|
|
|
|
710
|
1 |
|
class SubsetPooler(BaseEstimator, TransformerMixin): |
|
711
|
|
|
"""Wrapper for pool_duplicate_subsets(). Allows pool_duplicate_subsets() to be \ |
|
712
|
|
|
put into a pipeline with similar functions (e.g. using DataCleaner() or \ |
|
713
|
|
|
MVColHandler()). |
|
714
|
|
|
|
|
715
|
|
|
Parameters |
|
716
|
|
|
---------- |
|
717
|
|
|
col_dupl_ratio: float, default 0.2 |
|
718
|
|
|
Columns with a ratio of duplicates higher than "col_dupl_ratio" are considered \ |
|
719
|
|
|
in the further analysis. Columns with a lower ratio are not considered for \ |
|
720
|
|
|
pooling. |
|
721
|
|
|
dupl_thresh: float, default 0.2 |
|
722
|
|
|
The first subset with a duplicate threshold higher than "dupl_thresh" is \ |
|
723
|
|
|
chosen and aggregated. If no subset reaches the threshold, the algorithm \ |
|
724
|
|
|
continues with continuously smaller subsets until "min_col_pool" is reached. |
|
725
|
|
|
min_col_pool: integer, default 3 |
|
726
|
|
|
Minimum number of columns to pool. The algorithm attempts to combine as many \ |
|
727
|
|
|
columns as possible to suitable subsets and stops when "min_col_pool" is \ |
|
728
|
|
|
reached. |
|
729
|
|
|
return_details: bool, default False |
|
730
|
|
|
Provdies flexibility to return intermediary results. |
|
731
|
|
|
|
|
732
|
|
|
Returns |
|
733
|
|
|
------- |
|
734
|
|
|
data: pd.DataFrame |
|
735
|
|
|
""" |
|
736
|
|
|
|
|
737
|
1 |
|
def __init__( |
|
738
|
|
|
self, |
|
739
|
|
|
col_dupl_thresh=0.2, |
|
740
|
|
|
subset_thresh=0.2, |
|
741
|
|
|
min_col_pool=3, |
|
742
|
|
|
return_details=True, |
|
743
|
|
|
): |
|
744
|
|
|
self.col_dupl_thresh = col_dupl_thresh |
|
745
|
|
|
self.subset_thresh = subset_thresh |
|
746
|
|
|
self.min_col_pool = min_col_pool |
|
747
|
|
|
self.return_details = return_details |
|
748
|
|
|
|
|
749
|
1 |
|
def fit(self, data, target=None): |
|
750
|
|
|
return self |
|
751
|
|
|
|
|
752
|
1 |
|
@staticmethod |
|
753
|
1 |
|
def transform(data, target=None): |
|
754
|
|
|
data, subset_cols = pool_duplicate_subsets( |
|
755
|
|
|
data, |
|
756
|
|
|
col_dupl_thresh=0.2, |
|
757
|
|
|
subset_thresh=0.2, |
|
758
|
|
|
min_col_pool=3, |
|
759
|
|
|
return_details=True, |
|
760
|
|
|
) |
|
761
|
|
|
|
|
762
|
|
|
print("Combined columns:", len(subset_cols), subset_cols) |
|
763
|
|
|
|
|
764
|
|
|
return data |
|
765
|
|
|
|
akanz1 /
klib
