-
Notifications
You must be signed in to change notification settings - Fork 1
/
sd4py_extra.py
1215 lines (833 loc) · 50.5 KB
/
sd4py_extra.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
import numpy as np
import pandas as pd
import re
import seaborn as sns
import matplotlib.pyplot as plt
import matplotlib.colors as mcolors
from textwrap import wrap
import datetime
import networkx as nx
import sd4py
def bootstrapping(subgroups, data, metric_function, aggregation_function=None, ignore_defaults=False, number_simulations=100, frac=1/3, replace=True):
'''
Provides some estimate of variability for subgroups. Multiple samples (with replacement) are drawn from the data,
and subgroups are evaluated for each sample (using the metric_function).
The aggregation function is then applied to this data (e.g. to select 0.05 and 0.95 quantiles) to give a final description for each subgroup.
Parameters
----------------
subgroups: PySubgroup object or list of PySubgroup objects
The subgroup(s) for which to perform bootstrapping.
data: DataFrame
The data to be used to evaluate the subgroups; bootstrapping works by drawing samples from this data using replacement.
metric_function: function
The function to use to evaluate the how well the subgroup is working on an individual sample. Must have the following parameters: (sample, subgroup_sample), where `sample` is a sample of the data, and `subgroup_sample` is the same but filtered to only include subgroup members.
aggregation_function: function, optional
Used to aggregate across the samples. If not provided, the full list of scores (calculated by metric_function) over all the samples for each subgroup will be returned.
ignore_defaults: boolean, optional
If True, then the first row in data will be treated as containing 'default values' to be ignored in the processing.
number_simulations: int, optional
The number of samples to use.
frac: float, optional
The size of each sample as a proportion of the length of data. Can be reduced to decrease computational cost.
replace: boolean, optional
Set to False to override the default sampling with replacement strategy.
Returns
-----------
results
A dict with subgroup names as keys and bootstrapping results as values, or, when using only one subgroup, just the bootstrapping results.
aggregation
A dict with subgroup names as keys and aggregated results as values (or an empty dict if there is no aggregation function), or when using only one subgroup, just the aggregation (or None if there is no aggregation function).
'''
## This code is quite ugly but that was needed to speed things up.
samples = []
for x in range(number_simulations):
sample = data.sample(frac=frac, replace=False)
if replace:
sample = sample.sample(frac=1, replace=True)
if ignore_defaults:
sample = sample.replace(data.iloc[0,:], np.NaN)
samples.append(sample.reset_index(drop=True))
selectors = set()
for subgroup in subgroups:
for sel in subgroup.selectors:
selectors.add(sel)
selectors = list(selectors)
if isinstance(subgroups, sd4py.PySubgroupResults):
subgroups = subgroups.subgroups
if isinstance(subgroups, list):
def sample_indices(sample):
def get_indices(sel):
logical_indices = np.ones(sample.index.shape, dtype=bool)
if isinstance(sel, sd4py.PyNumericSelector):
if sel.include_lower_bound and sel.lower_bound != float("-inf"):
np.logical_and(logical_indices, sample[sel.attribute].values >= sel.lower_bound, out = logical_indices) ## It's about x10 faster to use .values (i.e. numpy arrays and therefore numpy functions)
elif sel.lower_bound != float("-inf"):
np.logical_and(logical_indices, sample[sel.attribute].values > sel.lower_bound, out = logical_indices)
if sel.include_upper_bound and sel.upper_bound != float("inf"):
np.logical_and(logical_indices, sample[sel.attribute].values <= sel.upper_bound, out = logical_indices)
elif sel.upper_bound != float("inf"):
np.logical_and(logical_indices, sample[sel.attribute].values < sel.upper_bound, out = logical_indices)
if isinstance(sel, sd4py.PyNominalSelector):
np.logical_and(logical_indices, sample[sel.attribute].astype(str).values == sel.value, out = logical_indices)
return logical_indices
return dict(zip(map(str, selectors) , map(get_indices, selectors)))
samples_indices = dict(zip(range(number_simulations), map(sample_indices, samples)))
def process_subgroup(subgroup):
def get_metric_values(args):
idx, sample = args
logical_indices = np.ones(sample.index.shape, dtype=bool)
for sel in subgroup.selectors:
np.logical_and(logical_indices, samples_indices[idx][str(sel)], out = logical_indices)
subgroup_sample = sample[logical_indices]
return metric_function(sample, subgroup_sample)
return list(map(get_metric_values, enumerate(samples)))
results = dict(zip(map(str, subgroups), map(process_subgroup, subgroups)))
if aggregation_function is not None:
aggregation = {key: aggregation_function(val) for key, val in results.items()}
return results, aggregation
else: ## Not a list
metric_values = []
for sample in samples:
subgroup_sample = subgroup.get_rows(sample)
metric_values.append(metric_function(sample, subgroup_sample))
if aggregation_function is None:
return metric_values, None
else:
return metric_values, aggregation_function(metric_values)
def confidence_intervals(subgroups, data, ignore_defaults=False, number_simulations=100, frac=1/3, replace=True):
'''
Provides some estimate of variability of the target value for subgroups. Uses bootstrapping to achieve this.
The target value and the size of each subgroup is calculated across 100 samples of the data. The 0.05 and 0.95 quantiles are returned per subgroup.
For numeric target variables, the mean within subgroup members is used; for nominal targets, the proportion of subgroup members belonging to the 'positive' class is used.
Parameters
----------------
subgroups: list of PySubgroup objects
The subgroup(s) for which to estimate confidence intervals.
data: DataFrame
The data to be used to evaluate the subgroups; bootstrapping works by drawing samples from this data using replacement.
ignore_defaults: boolean, optional
If True, then the first row in data will be treated as containing 'default values' to be ignored in the processing.
number_simulations: int, optional
The number of samples to use.
frac: float, optional
The size of each sample as a proportion of the length of data. Can be reduced to decrease computational cost.
replace: boolean, optional
Set to False to override the default sampling with replacement strategy.
Returns
-----------
bootstrapping_results: dict
A dict with subgroup names as keys and bootstrapping results as values, or, when using only one subgroup, just the bootstrapping results.
confidence_intervals: DataFrame
A DataFrame with the estimated confidence intervals, indexed by subgroup name.
'''
target = subgroups.target
if data.loc[:,target].dtype == 'object' or data.loc[:,target].dtype == 'bool' or data.loc[:,target].dtype.name == 'category': ## if nominal
def metric_function(sample, subgroup_sample):
subgroup_sample = subgroup_sample.loc[:,target]
sample = sample.loc[:,target]
population_share = subgroup_sample.count() / sample.count()
target_proportion = subgroup_sample.eq(subgroups.target_value).sum() / subgroup_sample.count() ## what proportion of values is equal to the target value
return population_share, target_proportion
else: ## if numeric
def metric_function(sample, subgroup_sample):
subgroup_sample = subgroup_sample.loc[:,target]
sample = sample.loc[:,target]
population_share = subgroup_sample.count() / sample.count()
average = subgroup_sample.mean()
return population_share, average
def aggregation_function(subgroup_values):
out = {
'proportion_lower': np.nanquantile([val[0] for val in subgroup_values], 0.05),
'proportion_upper': np.nanquantile([val[0] for val in subgroup_values], 0.95),
'target_lower': np.nanquantile([val[1] for val in subgroup_values], 0.05),
'target_upper': np.nanquantile([val[1] for val in subgroup_values], 0.95)
}
return out
bootstrapping_results, confidence_intervals = bootstrapping(subgroups, data, metric_function, aggregation_function,
ignore_defaults=ignore_defaults, number_simulations=number_simulations, frac=frac, replace=replace)
return bootstrapping_results, pd.DataFrame({'pattern':str(subgroup), **values} for subgroup, values in confidence_intervals.items())
def confidence_intervals_to_boxplots(bootstrapping_results_list, labels):
'''
Takes the outputs of the `confidence_intervals` function and creates a boxplot showing the distribution of the target value,
with the width of boxes indicating the relative sizes of the subgroups on average.
Parameters
----------------
bootstrapping_results_list: list
A list with subgroup bootstrapping results as values.
labels: list
The label to use for each subgroup.
Returns
-----------
fig: Figure
The matplotlib Figure of the boxplots
'''
averages = np.stack([np.array(x)[:,1] for x in bootstrapping_results_list])
for idx, row in enumerate(averages):
averages[idx][np.isnan(row)] = row[~np.isnan(row)].mean() # remove nan
widths = [np.array(x)[:,0].mean() for x in bootstrapping_results_list]
widths = 0.9 * np.array(widths) / np.max(widths) ## Box thickness relative to the maximum shown. Adjusted by 0.9 to avoid overlap
plt.boxplot(averages.T, vert=False, widths=widths, labels=labels)
plt.gca().xaxis.grid(True, linestyle='--')
return plt.gcf()
def confidence_precision_recall_f1(subgroups, data, ignore_defaults=False, number_simulations=100, frac=1/3, replace=True):
'''
Used to provide an estimate of how variable the performance of each subgroup is.
Applies to nominal variables, where the precision, recall and $F_1$ score are used to quantify how well a subgroup performs.
Parameters
----------------
subgroups: list of PySubgroup objects
The subgroup(s) for which to estimate confidence intervals.
data: DataFrame
The data to be used to evaluate the subgroups; bootstrapping works by drawing samples from this data using replacement.
ignore_defaults: boolean, optional
If True, then the first row in data will be treated as containing 'default values' to be ignored in the processing.
number_simulations: int, optional
The number of samples to use.
frac: float, optional
The size of each sample as a proportion of the length of data. Can be reduced to decrease computational cost.
replace: boolean, optional
Set to False to override the default sampling with replacement strategy.
Returns
-----------
bootstrapping_results: dict
A dict with subgroup names as keys and bootstrapping results as values, or, when using only one subgroup, just the bootstrapping results.
precision_recall_f1: DataFrame
A DataFrame with the estimated confidence intervals (0.05 and 0.95 quantiles from bootstrapping) on each of precision, recall and $F_1$, indexed by subgroup name.
'''
target = subgroups.target
target_value = subgroups.target_value
def metric_function(sample, subgroup_sample):
subgroup_sample = subgroup_sample.loc[:,target]
sample = sample.loc[:,target]
precision = subgroup_sample.values.__eq__(target_value).sum() / subgroup_sample.count() ## Use numpy arrays to check for equality since they're much faster
recall = subgroup_sample.values.__eq__(target_value).sum() / sample.values.__eq__(target_value).sum() ## Use numpy arrays to check for equality since they're much faster
f1 = (2 * precision * recall) / (precision + recall)
return precision, recall, f1
def aggregation_function(subgroup_values):
out = {
'precision_lower': np.nanquantile([val[0] for val in subgroup_values], 0.05),
'precision_upper': np.nanquantile([val[0] for val in subgroup_values], 0.95),
'recall_lower': np.nanquantile([val[1] for val in subgroup_values], 0.05),
'recall_upper': np.nanquantile([val[1] for val in subgroup_values], 0.95),
'f1_lower': np.nanquantile([val[2] for val in subgroup_values], 0.05),
'f1_upper': np.nanquantile([val[2] for val in subgroup_values], 0.95)
}
return out
bootstrapping_results, aggregation = bootstrapping(subgroups, data, metric_function, aggregation_function,
ignore_defaults=ignore_defaults, number_simulations=number_simulations, frac=frac, replace=replace)
return bootstrapping_results, pd.DataFrame({'pattern':str(subgroup), **values} for subgroup, values in aggregation.items())
def corrected_hedges_g(sample1, sample2):
'''
Estimates the effect size between two samples of a numeric variable.
This is the corrected Hedge's G; see <https://www.itl.nist.gov/div898/software/dataplot/refman2/auxillar/hedgeg.htm>.
Parameters
----------------
sample1: array
The first sample of values.
sample2: array
The second sample of values.
Returns
-----------
corrected_hedges_g: float
The estimated effect size.
'''
n_1 = sample1.count()
n_2 = sample2.count()
pooled_sd = np.sqrt((((n_1-1) * sample1.var()) + ((n_2-1) * sample2.var())) / (n_1 + n_2 - 2))
n = n_1 + n_2
bias_correction = ((n-3)/(n-2.25)) * np.sqrt((n - 2) / n)
return bias_correction * (sample1.mean() - sample2.mean()) / pooled_sd
def confidence_hedges_g(subgroups, data, ignore_defaults=False, number_simulations=100, frac=1/3, replace=True):
'''
Used to provide an estimate of the effect size for different subgroups when the target variable is numeric.
Parameters
----------------
subgroups: list of PySubgroup objects
The subgroup(s) for which to estimate confidence intervals.
data: DataFrame
The data to be used to evaluate the subgroups; bootstrapping works by drawing samples from this data using replacement.
target: string
The name of the target variable.
value: object, optional
For nominal target variables only. The value of the target variable that counts as the 'positive' class.
ignore_defaults: boolean, optional
If True, then the first row in data will be treated as containing 'default values' to be ignored in the processing.
number_simulations: int, optional
The number of samples to use.
frac: float, optional
The size of each sample as a proportion of the length of data. Can be reduced to decrease computational cost.
replace: boolean, optional
Set to False to override the default sampling with replacement strategy.
Returns
-----------
bootstrapping_results: dict
A dict with subgroup names as keys and bootstrapping results as values, or, when using only one subgroup, just the bootstrapping results.
confidence_hedges_g: DataFrame
A DataFrame with the estimated confidence intervals (0.05 and 0.95 quantiles from bootstrapping) on the effect size, indexed by subgroup name.
'''
target = subgroups.target
def metric_function(sample, subgroup_sample):
subgroup_sample = subgroup_sample.loc[:,target]
sample = sample.loc[:,target]
complement = sample[~sample.index.isin(subgroup_sample.index)]
proportion = subgroup_sample.count() / sample.count()
hedges_g = corrected_hedges_g(subgroup_sample, complement)
return proportion, hedges_g
def aggregation_function(subgroup_values):
out = {
'proportion_lower': np.nanquantile([val[0] for val in subgroup_values], 0.05),
'proportion_upper': np.nanquantile([val[0] for val in subgroup_values], 0.95),
'hedges_g_lower': np.nanquantile([val[1] for val in subgroup_values], 0.05),
'hedges_g_upper': np.nanquantile([val[1] for val in subgroup_values], 0.95)
}
return out
bootstrapping_results, aggregation = bootstrapping(subgroups, data, metric_function, aggregation_function,
ignore_defaults=ignore_defaults, number_simulations=number_simulations, frac=frac, replace=replace)
return bootstrapping_results, pd.DataFrame({'pattern':str(subgroup), **values} for subgroup, values in aggregation.items())
def odds_ratio_ci(sample1, sample2):
'''
Estimates the effect size between two samples of a binary nominal variable.
This is the odds ratio, which allows us to estimate confidence intervals directly from the confusion matrix.
Parameters
----------------
sample1: array
The first sample of values.
sample2: array
The second sample of values.
Returns
-----------
odds_ratio: float
The estimated effect size.
lower: float
Lower confidence interval on the estimated effect size.
upper: float
Upper confidence interval on the estimated effect size.
'''
a = sample1.eq(True).sum() # subgroup == True and column == value
b = sample1.eq(False).sum() # subgroup == True and column != value
c = sample2.eq(True).sum() # subgroup == False and column == value
d = sample2.eq(False).sum() # subgroup == False and column != value
if min(a,b,c,d) == 0:
return np.NaN, np.NaN, np.NaN
odds_ratio = (a * d) / (b * c)
lower = np.exp(np.log(odds_ratio) - (1.96 * np.sqrt((1/a) + (1/b) + (1/c) + (1/d))))
upper = np.exp(np.log(odds_ratio) + (1.96 * np.sqrt((1/a) + (1/b) + (1/c) + (1/d))))
return odds_ratio, lower, upper
def find_interesting_columns(subgroup, data, use_complement = True, ignore_defaults = False, columns_to_ignore=[]):
'''
Makes it easier to find 'interesting' columns for particular subgroup by returning the estimated effect size for each variable in the dataset
(i.e., if a variable has a large effect size then the subgroup is extreme with respect to that variable).
Provides both interesting numeric and interesting nominal columns.
Corrected Hedge's G is used to estimate effect size on numeric variables, and the odds ratio (and its confidence intervals) is used for nominal variables.
Parameters
----------------
subgroup: PySubgroup
The subgroup for which to find interesting columns.
data: DataFrame
The data to be used to evaluate the subgroups; bootstrapping works by drawing samples from this data using replacement.
use_complement: boolean, optional
If True, subgroup members will be compared to non-subgroup members. Otherwise, subgroup members will be compared to the full dataset (including subgroup members).
ignore_defaults: boolean, optional
If True, then the first row in data will be treated as containing 'default values' to be ignored in the processing.
columns_to_ignore: list, optional
A list of columns to ignore, for example these could be the target variable and/or selector variables since they are already known to be interesting.
Returns
-----------
numeric_columns: dict
A dictionary with variable names as keys and estimated effect sizes as values.
nominal_columns: dict
A dictionary with variable names as keys and estimated effect sizes as values, for nominal variables.
'''
numeric_columns = {}
nominal_columns = {}
subgroup_indices = subgroup.get_indices(data)
if ignore_defaults:
data = data.replace(data.iloc[0,:], np.NaN)
for column in data:
if column in columns_to_ignore:
continue
column = data[column]
if (np.issubdtype(column.dtype, np.datetime64) or np.issubdtype(column.dtype, np.timedelta64)): ## these need to be converted to a straightforward numeric format
column = pd.to_numeric(column)
column = (column - column.mean()) / column.std() ## Just to get to a reasonable timescale, otherwise it's nanoseconds or something like that
subgroup_rows = column.loc[subgroup_indices]
if use_complement:
population_rows = column.drop(subgroup_indices, axis=0)
else:
population_rows = column
if column.dtype == 'object' or column.dtype == 'bool' or column.dtype.name == 'category': ## nominals
vals, counts = np.unique(column, return_counts=True)
for value in vals[np.argsort(-counts)][:5]: ## 5 most common values for this variable; each providing a feature-value pair to investigate
nominal_columns[(column.name, value)] = odds_ratio_ci(subgroup_rows == value, population_rows == value)
else: ## numerics
numeric_columns[column.name] = corrected_hedges_g(subgroup_rows, population_rows)
return numeric_columns, nominal_columns
def most_interesting_columns(subgroup, data, columns_to_ignore=[]):
'''
To support visualisation of a single subgroup, uses the `find_interesting_columns` function to pick the 10 most numeric and 10 most interesting nominal values for a subgroup.
Corrected Hedge's G is used to estimate effect size on numeric variables, and lower confidence on the odds ratio is used for nominal variables.
Parameters
----------------
subgroup: PySubgroup
The subgroup for which to find interesting columns.
data: DataFrame
The data to be used to evaluate the subgroups; bootstrapping works by drawing samples from this data using replacement.
columns_to_ignore: list, optional
A list of columns to ignore, for example these could be the target variable and/or selector variables since they are already known to be interesting.
Returns
-----------
most_interesting_numeric: DataFrame
A pandas DataFrame with variable names as index and estimated effect sizes as values.
most_interesting_nominal: DataFrame
A pandas DataFrame with variable names as index and estimated effect sizes as values, for nominal variables.
'''
interesting_numeric, interesting_nominal = find_interesting_columns(subgroup, data, columns_to_ignore=columns_to_ignore)
interesting_numeric = pd.DataFrame(interesting_numeric.values(), index=interesting_numeric.keys())
interesting_numeric = interesting_numeric.dropna()
interesting_nominal = pd.DataFrame(interesting_nominal.values(), index=interesting_nominal.keys())
interesting_nominal = interesting_nominal.dropna()
if len(interesting_numeric) > 0:
interesting_numeric = interesting_numeric.dropna()
most_interesting_numeric = interesting_numeric.iloc[interesting_numeric[0].abs().argsort()][::-1][0].iloc[:10] ## Find the 10 most interesting by effect size
else:
most_interesting_numeric = interesting_numeric
if len(interesting_nominal) > 0:
max_lower = interesting_nominal.loc[interesting_nominal[1].abs().groupby(level=0).idxmax()][1] ## Maximum lower confidence interval
max_lower = max_lower.iloc[max_lower.abs().argsort()][::-1]
max_lower
min_upper = interesting_nominal.loc[interesting_nominal[2].abs().groupby(level=0).idxmin()][2]
min_upper = min_upper[interesting_nominal.groupby(level=0).count()[2].values > 2]
min_upper = min_upper[min_upper > 0]
min_upper = (1 / min_upper)
min_upper = min_upper.iloc[min_upper.abs().argsort()][::-1]
min_upper
most_interesting_nominal = pd.concat([max_lower, min_upper]).sort_values(ascending=False).iloc[:10] ## 10 most interesting by having especially high or especially low odds ratio
else:
most_interesting_nominal = interesting_nominal
return most_interesting_numeric, most_interesting_nominal
def radar_plot(data, prop_scale=3, subplot=111, text_size = 10, axis_padding = 15, ymins = None, ymaxes = None):
'''
Creates a custom radar plot, where axis names and units can vary. Note that radar plots are poorly-supported by matplotlib and things like tight_layout will not work.
Parameters
----------------
data: DataFrame
A dataframe where columns are variables and rows are the groups (each group will become a polygon).
prop_scale: float, optional
Used to control the size of the innermost circle (where axes begin) compared to the rest of the plot.
subplot: int, optional
Used to determine which subplot to draw the radar plot onto.
text_size: int, optional
Used to modify the text size of axis labels.
axis_padding: int, optional
Used to modify the padding around axis names (to prevent them overlapping with axis tick labels). Modified by position, so more horizontal axes get more padding (since the names are more likely to overlap with the ticks).
ymins: list, optional
Used to set the beginning of each axis
ymaxes: list, optional
Used to set the end of each axis
Returns
-----------
ax: Axis
The matplotlib Axis of the radar plot
'''
num_variables = len(data.columns) ## Number of columns/variables
if num_variables < 3:
num_variables = 3 ## So that we always have a shape with an area
angles = [n / float(num_variables) * 2 * np.pi for n in range(num_variables)]
angles += angles[:1] # And back to the first position
# Initialise the radar plot
ax = plt.subplot(subplot, polar=True)
# To put the first axis on top:
ax.set_theta_offset(np.pi / 2)
ax.set_theta_direction(-1)
# Draw one axis per variable + add names
x_ticks = plt.xticks(angles[:len(data.columns)], data.columns, size=text_size+1)
ax.spines['polar'].set_color('grey')
ax.yaxis.set_visible(False) # Axes and appropriate scales will be drawn later, using polar_twin
ax.grid(False) # Axes and appropriate scales will be drawn later, using polar_twin
#ax.tick_params(axis='x', which='major', pad=axis_padding) #Space the axis labels a bit
for idx, x_tick in enumerate(x_ticks[0]):
x_tick.set_pad((axis_padding//5) + (axis_padding * np.abs(np.sin(angles[idx]) ** 2))) ## This seems to give slightly better padding than the previous attempt
def polar_twin(ax, ymin, ymax, angle=20): # function to make a new axis with appropriate tick marks
ax2 = ax.figure.add_axes(ax.get_position(), projection='polar',
label='twin', frameon=False,
theta_direction=ax.get_theta_direction(),
theta_offset=ax.get_theta_offset())
ax2.xaxis.set_visible(False)
#labels = [ymin, ymax]
labels = [ymin, ymin + ((ymax-ymin) * 1/3), ymin + ((ymax-ymin) * 2/3), ymax]
if hasattr(ymin, 'strftime'):
labels = [item.strftime('%Y-%m-%d\n%H:%M:%S') for item in labels]
else:
try:
labels = ["{:.2f}".format(float(item)) for item in labels]
except:
labels = ["{0} days\n{1:02d}:{2:02d}:{3:02d}".format(*item.components) for item in labels]
ax2.set_ylim(0, 1+prop_scale)
ax2.set_rgrids([1,1+(prop_scale/3),1+(2*prop_scale/3),1+prop_scale], labels, angle, size=text_size, ha="center", va="center")
# To ensure that the original axes tick labels are on top of
# whatever is plotted in the twinned axes. Tick labels will be drawn twice.
for label in ax.get_yticklabels():
ax.figure.texts.append(label)
ax2.grid(False)
return ax2
if ymins is None:
ymins = data.min()
if ymaxes is None:
ymaxes = data.max()
for idx, colname in enumerate(data):
col = data.loc[:,colname]
ymin = ymins[colname]
ymax = ymaxes[colname]
angle = idx * 360 / num_variables
ax_latest = polar_twin(ax, ymin, ymax, angle)
ax_latest.set_zorder(100) # so axis grid doesn't appear in front of other content
ax_latest.grid(True)
ax_latest.set_zorder(10) # so axis grid doesn't appear in front of other content
def plot_polygon(row, angles, colour, label, linestyle):
# Draws the polygon for one subgroup onto the radar plot
values=row.flatten().tolist()
if len(values) < 3:
values += np.ones(3 - len(values)).tolist()
values += values[:1] ## To go back to the start
ax.set_ylim(0, 1+prop_scale)
ax.plot(angles, values, linewidth=2, linestyle=linestyle, color=colour, label=label)
ax.fill(angles, values, colour, alpha=0.1)
data_norm = 1 + (prop_scale * (data - ymins) / (ymaxes - ymins)) # Scale the data to match to the labels
for idx, row in enumerate(data_norm.values):
# We use standard 'tableau' colours from matplotlib, and varying linestyle
plot_polygon(row, angles, list(mcolors.TABLEAU_COLORS)[idx % len(mcolors.TABLEAU_COLORS)],
label=str(data_norm.index[idx]), linestyle=['solid','dashed','dotted','dashdot'][idx%4])
# Draw a legend now that the polygons have been plotted
ax.legend(loc='upper right', bbox_to_anchor=(0.1, 0.1))
ax.set_zorder(100) # so axis grid doesn't appear in front of other content
ax.patch.set_visible(False) # so axis grid doesn't appear in front of other content
return ax
def subgroup_overview(subgroup, selection_data, visualisation_data=None, use_complement=True, axis_padding = 15):
'''
Creates a four-panel matplotlib visualisation for a single subgroup.
From left to right, top to bottom, this shows:
(i) the distribution of target values for the subgroup and its complement,
(ii) the selector variable average values,
(iii) average values for other numeric variables, and
(iv) relative frequency of certain (variable, value) pairs for other nominal variables.}
Note that radar plots are poorly-supported by matplotlib and things like tight_layout will not work.
Parameters
----------------
subgroup: PySubgroup
The subgroup to be visualised.
selection_data: DataFrame
The subgroup will be applied to this data, to select subgroup members. From this, the most interesting columns to visualise will be chosen. If visualisation_data is not provided, this will also be the data used to compute the values that are visualised.
visualisation_data: DataFrame
If desired, a second dataset can be used to provide the data that is visualised (but not used to select the 'most interesting columns').
use_complement: boolean, optional
If True, subgroup members will be compared to non-subgroup members. Otherwise, subgroup members will be compared to the full dataset (including subgroup members).
Returns
-----------
fig: Figure
The matplotlib Figure of the subgroup overview.
'''
target = subgroup.target
if visualisation_data is None:
visualisation_data = selection_data
def visualise_columns(numeric_columns=None, nominal_columns=None, nominal_values=None, prop_scale=2.5, subplot=111):
## This function finds appropriate ymins and ymaxes for plotting each axis, and then calls the radar_plot function
means = pd.DataFrame()
proportions = pd.DataFrame()
numeric_ymins = pd.Series(dtype=object)
numeric_ymaxes = pd.Series(dtype=object)
nominal_ymins = pd.Series(dtype=object)
nominal_ymaxes = pd.Series(dtype=object)
subgroup_indices = subgroup.get_indices(visualisation_data)
## Numerics
if numeric_columns is not None:
subgroup_means = visualisation_data.loc[subgroup_indices][numeric_columns].mean(numeric_only=False)
if use_complement:
means = pd.concat([
subgroup_means,
visualisation_data.drop(subgroup_indices, axis=0)[numeric_columns].mean(numeric_only=False)
], axis=1).T.set_index([['Subgroup', 'Complement']])
else:
means = pd.concat([
subgroup_means,
visualisation_data[numeric_columns].mean(numeric_only=False)
], axis=1).T.set_index([['Subgroup', 'Population']])
vis_data_numerics = visualisation_data[numeric_columns]
numeric_ymins = vis_data_numerics.mean(numeric_only=False) - (vis_data_numerics.std(numeric_only=False))
numeric_ymins = pd.concat([
numeric_ymins,
subgroup_means
],axis=1).T.min(numeric_only=False) ## Minimum of (complement - 1 std) and (subgroup_mean)
numeric_ymaxes = vis_data_numerics.mean(numeric_only=False) + (vis_data_numerics.std(numeric_only=False))
numeric_ymaxes = pd.concat([
numeric_ymaxes,
subgroup_means
],axis=1).T.max(numeric_only=False) ## Maximum of (complement + 1 std) and (subgroup_mean)
## Now the nominals
if nominal_columns is not None:
nominal_data = visualisation_data.loc[:,nominal_columns].astype(str)
subgroup_proportions = nominal_data.loc[subgroup_indices, :].eq(nominal_values).sum() / nominal_data.loc[subgroup_indices, :].count()
if use_complement:
proportions = pd.concat([
subgroup_proportions,
nominal_data.drop(subgroup_indices, axis=0).eq(nominal_values).sum() / nominal_data.drop(subgroup_indices, axis=0).count()
], axis=1).T.set_index([['Subgroup', 'Complement']])
else:
proportions = pd.concat([
subgroup_proportions,
nominal_data.eq(nominal_values).sum() / nominal_data.count()
], axis=1).T.set_index([['Subgroup', 'Population']])
nominal_ymins = ((2* proportions) - 1).min() ## same distance below the proportion as above it (up to 1), minimum across subgroup and complement
nominal_ymins = pd.concat([nominal_ymins, pd.Series(0, index=nominal_ymins.index)], axis=1).T.max() ## set to zero if currently below zero
nominal_ymaxes = (2* proportions).max() ## same distance above the proportion as below it (down to 1), maximum across subgroup and complement
nominal_ymaxes = pd.concat([nominal_ymaxes, pd.Series(1, index=nominal_ymaxes.index)], axis=1).T.min() ## set to 1 if currently above 1
nominal_ymins.index = ["{0} == {1}".format(*x) for x in zip(nominal_columns, nominal_values)]
nominal_ymaxes.index = ["{0} == {1}".format(*x) for x in zip(nominal_columns, nominal_values)]
proportions.columns = ["{0} == {1}".format(*x) for x in zip(nominal_columns, nominal_values)]
total = pd.concat([means, proportions], axis=1)
ymins = pd.concat([numeric_ymins, nominal_ymins])
ymaxes = pd.concat([numeric_ymaxes, nominal_ymaxes])
return radar_plot(total, prop_scale=prop_scale, ymins=ymins, ymaxes=ymaxes, subplot=subplot, axis_padding=axis_padding)
## Target
ax = plt.subplot(221)
subgroup_indices = subgroup.get_indices(visualisation_data)
if visualisation_data[target].dtype == 'object' or visualisation_data[target].dtype == 'bool' or visualisation_data[target].dtype.name == 'category':
## For nominal target, use a stacked barchart to visualise distribution
if use_complement:
pd.concat([
pd.Series(*np.unique(visualisation_data.loc[subgroup_indices][target], return_counts=True)[::-1], name='Subgroup') \
/ visualisation_data.loc[subgroup_indices][target].count(),
pd.Series(*np.unique(visualisation_data.drop(subgroup_indices, axis=0)[target], return_counts=True)[::-1], name='Complement') \
/ visualisation_data.drop(subgroup_indices, axis=0)[target].count()
],axis=1).T.plot(kind='barh', stacked=True, cmap=plt.get_cmap('Set2'), ax=ax)
for container in ax.containers:
ax.bar_label(container, label_type='center', fmt="%.2f")
ax.legend()
else:
pd.concat([
pd.Series(*np.unique(visualisation_data.loc[subgroup_indices][target], return_counts=True)[::-1], name='Subgroup') \
/ visualisation_data.loc[subgroup_indices][target].count(),
pd.Series(*np.unique(visualisation_data[target], return_counts=True)[::-1], name='Complement') \
/ visualisation_data[target].count()
],axis=1).T.plot(kind='barh', stacked=True, cmap=plt.get_cmap('Set2'), ax=ax)
for container in ax.containers:
ax.bar_label(container, label_type='center', fmt="%.2f")
ax.legend()
else:
## For numeric target, use an estimated probability density plot
if use_complement:
sns.kdeplot(visualisation_data.loc[subgroup_indices][target], linewidth=2, label='Subgroup')
sns.kdeplot(visualisation_data.drop(subgroup_indices, axis=0)[target], linewidth=2, label='Complement', linestyle='dashed')
ax.legend()
else:
sns.kdeplot(visualisation_data.loc[subgroup_indices][target], linewidth=2, label='Subgroup')
sns.kdeplot(visualisation_data[target], linewidth=2, label='Population', linestyle='dashed')
ax.legend()
ax.set_title('Target', pad =20)
## Selectors
numeric_selectors = []
nominal_selectors = []
nominal_selector_values = []
for selector in subgroup.selectors:
if isinstance(selector, sd4py.PyNumericSelector):
numeric_selectors.append(selector.attribute)
else:
nominal_selectors.append(selector.attribute)
nominal_selector_values.append(str(selector.value))
if len(numeric_selectors) == 0:
numeric_selectors = None
if len(nominal_selectors) == 0:
nominal_selectors = None
ax = visualise_columns(numeric_columns=numeric_selectors, nominal_columns=nominal_selectors, nominal_values=nominal_selector_values, subplot=222)
ax.set_title('Selectors', pad =20)
## Additional variables
columns_to_ignore = [s.attribute for s in subgroup.selectors] ## Selectors will already be visualised
columns_to_ignore += [target] ## Target will already be visualised
most_interesting_numeric, most_interesting_nominal = most_interesting_columns(subgroup, selection_data, columns_to_ignore=columns_to_ignore)
## Numeric
if len(most_interesting_numeric) > 0:
ax = visualise_columns(numeric_columns=most_interesting_numeric.index.tolist(), subplot=223)
ax.set_title('Additional Numeric Variables', pad =20)
## Nominals
if len(most_interesting_nominal) > 0:
columns = [x[0] for x in most_interesting_nominal.index]
values = [str(x[1]) for x in most_interesting_nominal.index]
ax = visualise_columns(nominal_columns=columns, nominal_values=values, subplot=224)
ax.set_title('Additional Nominal Variables', pad =20)
return plt.gcf()
def jaccard_visualisation(subgroups, data, minimum_jaccard=0, labels=None):
'''
Shows the similarity between a selection of subgroups. Uses the Jaccard similarity between each pair of subgroups to construct edges in a network diagram.
Parameters
----------------
subgroups: list of PySubgroup objects
The subgroups to visualise.
data: DataFrame
The data to be used to evaluate the similarity between pairs of subgroups.
minimum_jaccard: float
An edge will only be drawn between two subgroups if their Jaccard similarity is above this value.
labels: list
The label to use for each subgroup.
Returns
-----------
fig: Figure
The matplotlib Figure of the boxplots
'''
if labels is None:
labels = [str(sg) for sg in subgroups]
adjacency = np.zeros((len(subgroups), len(subgroups)))
for idx1, subgroup1 in enumerate(subgroups):
for idx2, subgroup2 in enumerate(subgroups):
if idx1 < idx2:
indices1 = subgroup1.get_indices(data)
indices2 = subgroup2.get_indices(data)
adjacency[idx1, idx2] = indices1.intersection(indices2).size / indices1.union(indices2).size
G = nx.from_numpy_matrix(adjacency * (adjacency > minimum_jaccard))
G = nx.relabel_nodes(G, mapping={idx:sg for idx, sg in enumerate(labels)})
pos = nx.spring_layout(G, seed=10) # seed so the results are consistent each time
# nodes
nx.draw_networkx_nodes(G, pos, node_size=500, alpha=0.5)
# edges
nx.draw_networkx_edges(
G, pos, alpha=0.2,
width = [7.5 * x for x in nx.get_edge_attributes(G,'weight').values()]
)
# labels
nx.draw_networkx_labels(G, pos, font_size=12, font_family="sans-serif")