Total Complexity | 49 |
Total Lines | 803 |
Duplicated Lines | 15.44 % |
Changes | 0 |
Duplicate code is one of the most pungent code smells. A rule that is often used is to re-structure code once it is duplicated in three or more places.
Common duplication problems, and corresponding solutions are:
Complex classes like responsibly.fairness.interventions.threshold often do a lot of different things. To break such a class down, we need to identify a cohesive component within that class. A common approach to find such a component is to look for fields/methods that share the same prefixes, or suffixes.
Once you have determined the fields that belong together, you can apply the Extract Class refactoring. If the component makes sense as a sub-class, Extract Subclass is also a candidate, and is often faster.
1 | """ |
||
2 | Post-processing fairness intervension by choosing thresholds. |
||
3 | |||
4 | There are multiple definitions for choosing the thresholds: |
||
5 | |||
6 | 1. Single threshold for all the sensitive attribute values |
||
7 | that minimizes cost. |
||
8 | 2. A threshold for each sensitive attribute value |
||
9 | that minimize cost. |
||
10 | 3. A threshold for each sensitive attribute value |
||
11 | that achieve independence and minimize cost. |
||
12 | 4. A threshold for each sensitive attribute value |
||
13 | that achieve equal FNR (equal opportunity) and minimize cost. |
||
14 | 5. A threshold for each sensitive attribute value |
||
15 | that achieve separation (equalized odds) and minimize cost. |
||
16 | |||
17 | The code is based on `fairmlbook repository <https://github.com/fairmlbook/fairmlbook.github.io>`_. |
||
18 | |||
19 | References: |
||
20 | - Hardt, M., Price, E., & Srebro, N. (2016). |
||
21 | `Equality of opportunity in supervised learning |
||
22 | <https://arxiv.org/abs/1610.02413>`_ |
||
23 | In Advances in neural information processing systems |
||
24 | (pp. 3315-3323). |
||
25 | - `Attacking discrimination with |
||
26 | smarter machine learning by Google |
||
27 | <https://research.google.com/bigpicture/attacking-discrimination-in-ml/>`_. |
||
28 | |||
29 | """ |
||
30 | |||
31 | # pylint: disable=no-name-in-module,ungrouped-imports |
||
32 | |||
33 | from collections import Counter |
||
34 | |||
35 | import matplotlib.pylab as plt |
||
36 | import numpy as np |
||
37 | import pandas as pd |
||
38 | import seaborn as sns |
||
39 | from matplotlib.ticker import AutoMinorLocator |
||
40 | from scipy.spatial import Delaunay |
||
41 | |||
42 | from responsibly.fairness.metrics.score import roc_curve_by_attr |
||
43 | from responsibly.fairness.metrics.utils import _groupby_y_x_sens |
||
44 | from responsibly.fairness.metrics.visualization import plot_roc_curves |
||
45 | |||
46 | |||
47 | TRINARY_SEARCH_TOL = 1e-3 |
||
48 | |||
49 | |||
50 | def _strictly_increasing(arr): |
||
51 | return (np.diff(arr) >= 0).all() |
||
52 | |||
53 | |||
54 | def _titlify(text): |
||
55 | text = text.replace('_', ' ').title() |
||
56 | if text == 'Fnr': |
||
57 | text = 'FNR' |
||
58 | return text |
||
59 | |||
60 | |||
61 | def _ternary_search_float(f, left, right, tol): |
||
62 | """Trinary search: minimize f(x) over [left, right], to within +/-tol in x. |
||
63 | |||
64 | Works assuming f is quasiconvex. |
||
65 | """ |
||
66 | |||
67 | while right - left > tol: |
||
68 | left_third = (2 * left + right) / 3 |
||
69 | right_third = (left + 2 * right) / 3 |
||
70 | if f(left_third) < f(right_third): |
||
71 | right = right_third |
||
72 | else: |
||
73 | left = left_third |
||
74 | return (right + left) / 2 |
||
75 | |||
76 | |||
77 | def _ternary_search_domain(f, domain): |
||
78 | """Trinary search: minimize f(x) over a domain (sequence). |
||
79 | |||
80 | Works assuming f is quasiconvex and domain is ascending sorted. |
||
81 | |||
82 | BUGGY, DO NOT USE |
||
83 | |||
84 | >>> arr = np.concatenate([np.arange(10, 2, -1), np.arange(2, 20)]) |
||
85 | >>> t1 = _ternary_search_domain(lambda t: arr[t], range(len(arr))) |
||
86 | >>> t2 = np.argmin(arr) |
||
87 | |||
88 | >>> assert t1 == t2 |
||
89 | >>> assert arr[t1] == arr[t2] |
||
90 | """ |
||
91 | |||
92 | left = 0 |
||
93 | right = len(domain) - 1 |
||
94 | changed = True |
||
95 | |||
96 | while changed and left != right: |
||
97 | |||
98 | changed = False |
||
99 | |||
100 | left_third = (2 * left + right) // 3 |
||
101 | right_third = (left + 2 * right) // 3 |
||
102 | |||
103 | if f(domain[left_third]) < f(domain[right_third]): |
||
104 | right = right_third - 1 |
||
105 | changed = True |
||
106 | else: |
||
107 | left = left_third + 1 |
||
108 | changed = True |
||
109 | |||
110 | return domain[(left + right) // 2] |
||
111 | |||
112 | |||
113 | def _cost_function(fpr, tpr, base_rate, cost_matrix): |
||
114 | """Compute the cost of given (fpr, tpr). |
||
115 | |||
116 | [[tn, fp], [fn, tp]] |
||
117 | """ |
||
118 | |||
119 | fp = fpr * (1 - base_rate) |
||
120 | tn = (1 - base_rate) - fp |
||
121 | tp = tpr * base_rate |
||
122 | fn = base_rate - tp |
||
123 | |||
124 | conf_matrix = np.array([tn, fp, fn, tp]) |
||
125 | |||
126 | return (conf_matrix * np.array(cost_matrix).ravel()).sum() |
||
127 | |||
128 | |||
129 | def _extract_threshold(roc_curves): |
||
130 | return next(iter(roc_curves.values()))[2] |
||
131 | |||
132 | |||
133 | def _first_index_above(arr, value): |
||
134 | """Find the smallest index i for which array[i] > value. |
||
135 | |||
136 | If no such value exists, return len(array). |
||
137 | """ |
||
138 | |||
139 | assert _strictly_increasing(arr), ( |
||
140 | 'arr should be stricktly increasing.') |
||
141 | |||
142 | arr = np.array(arr) |
||
143 | v = np.concatenate([arr > value, [1]]) |
||
144 | return np.argmax(v, axis=0) |
||
145 | |||
146 | |||
147 | def _calc_acceptance_rate(fpr, tpr, base_rate): |
||
148 | return (fpr * (1 - base_rate) |
||
149 | + tpr * base_rate) |
||
150 | |||
151 | |||
152 | def find_single_threshold(roc_curves, base_rates, proportions, |
||
153 | cost_matrix): |
||
154 | """Compute single threshold that minimizes cost. |
||
155 | |||
156 | :param roc_curves: Receiver operating characteristic (ROC) |
||
157 | by attribute. |
||
158 | :type roc_curves: dict |
||
159 | :param base_rates: Base rate by attribute. |
||
160 | :type base_rates: dict |
||
161 | :param proportions: Proportion of each attribute value. |
||
162 | :type proportions: dict |
||
163 | :param cost_matrix: Cost matrix by [[tn, fp], [fn, tp]]. |
||
164 | :type cost_matrix: sequence |
||
165 | :return: Threshold, FPR and TPR by attribute and cost value. |
||
166 | :rtype: tuple |
||
167 | |||
168 | """ |
||
169 | |||
170 | def total_cost_function(index): |
||
171 | total_cost = 0 |
||
172 | |||
173 | for group, roc in roc_curves.items(): |
||
174 | fpr = roc[0][index] |
||
175 | tpr = roc[1][index] |
||
176 | |||
177 | group_cost = _cost_function(fpr, tpr, |
||
178 | base_rates[group], cost_matrix) |
||
179 | |||
180 | group_cost *= proportions[group] |
||
181 | |||
182 | total_cost += group_cost |
||
183 | |||
184 | return -total_cost |
||
185 | |||
186 | thresholds = _extract_threshold(roc_curves) |
||
187 | |||
188 | cost_per_threshold = [total_cost_function(index) |
||
189 | for index in range(len(thresholds))] |
||
190 | cutoff_index = np.argmin(cost_per_threshold) |
||
191 | |||
192 | fpr_tpr = {group: (roc[0][cutoff_index], roc[1][cutoff_index]) |
||
193 | for group, roc in roc_curves.items()} |
||
194 | |||
195 | cost = total_cost_function(cutoff_index) |
||
196 | |||
197 | return thresholds[cutoff_index], fpr_tpr, cost |
||
198 | |||
199 | |||
200 | def find_min_cost_thresholds(roc_curves, base_rates, proportions, cost_matrix): |
||
201 | """Compute thresholds by attribute values that minimize cost. |
||
202 | |||
203 | :param roc_curves: Receiver operating characteristic (ROC) |
||
204 | by attribute. |
||
205 | :type roc_curves: dict |
||
206 | :param base_rates: Base rate by attribute. |
||
207 | :type base_rates: dict |
||
208 | :param proportions: Proportion of each attribute value. |
||
209 | :type proportions: dict |
||
210 | :param cost_matrix: Cost matrix by [[tn, fp], [fn, tp]]. |
||
211 | :type cost_matrix: sequence |
||
212 | :return: Thresholds, FPR and TPR by attribute and cost value. |
||
213 | :rtype: tuple |
||
214 | |||
215 | """ |
||
216 | # pylint: disable=cell-var-from-loop |
||
217 | |||
218 | cutoffs = {} |
||
219 | fpr_tpr = {} |
||
220 | |||
221 | cost = 0 |
||
222 | thresholds = _extract_threshold(roc_curves) |
||
223 | |||
224 | for group, roc in roc_curves.items(): |
||
225 | def group_cost_function(index): |
||
226 | fpr = roc[0][index] |
||
227 | tpr = roc[1][index] |
||
228 | return -_cost_function(fpr, tpr, |
||
229 | base_rates[group], cost_matrix) |
||
230 | |||
231 | cost_per_threshold = [group_cost_function(index) |
||
232 | for index in range(len(thresholds))] |
||
233 | cutoff_index = np.argmin(cost_per_threshold) |
||
234 | |||
235 | cutoffs[group] = thresholds[cutoff_index] |
||
236 | |||
237 | fpr_tpr[group] = (roc[0][cutoff_index], |
||
238 | roc[1][cutoff_index]) |
||
239 | |||
240 | cost += group_cost_function(cutoff_index) * proportions[group] |
||
241 | |||
242 | return cutoffs, fpr_tpr, cost |
||
243 | |||
244 | |||
245 | def get_acceptance_rate_indices(roc_curves, base_rates, |
||
246 | acceptance_rate_value): |
||
247 | indices = {} |
||
248 | |||
249 | for group, roc in roc_curves.items(): |
||
250 | # can be calculated outside the function |
||
251 | acceptance_rates = _calc_acceptance_rate(fpr=roc[0], |
||
252 | tpr=roc[1], |
||
253 | base_rate=base_rates[group]) |
||
254 | |||
255 | index = _first_index_above(acceptance_rates, |
||
256 | acceptance_rate_value) |
||
257 | |||
258 | indices[group] = index |
||
259 | |||
260 | return indices |
||
261 | |||
262 | |||
263 | View Code Duplication | def find_independence_thresholds(roc_curves, base_rates, proportions, |
|
264 | cost_matrix): |
||
265 | """Compute thresholds that achieve independence and minimize cost. |
||
266 | |||
267 | :param roc_curves: Receiver operating characteristic (ROC) |
||
268 | by attribute. |
||
269 | :type roc_curves: dict |
||
270 | :param base_rates: Base rate by attribute. |
||
271 | :type base_rates: dict |
||
272 | :param proportions: Proportion of each attribute value. |
||
273 | :type proportions: dict |
||
274 | :param cost_matrix: Cost matrix by [[tn, fp], [fn, tp]]. |
||
275 | :type cost_matrix: sequence |
||
276 | :return: Thresholds, FPR and TPR by attribute and cost value. |
||
277 | :rtype: tuple |
||
278 | |||
279 | """ |
||
280 | |||
281 | cutoffs = {} |
||
282 | |||
283 | def total_cost_function(acceptance_rate_value): |
||
284 | # todo: move demo here + multiple cost |
||
285 | # + refactor - use threshold to calculate |
||
286 | # acceptance_rate_value |
||
287 | indices = get_acceptance_rate_indices(roc_curves, base_rates, |
||
288 | acceptance_rate_value) |
||
289 | |||
290 | total_cost = 0 |
||
291 | |||
292 | for group, roc in roc_curves.items(): |
||
293 | index = indices[group] |
||
294 | |||
295 | fpr = roc[0][index] |
||
296 | tpr = roc[1][index] |
||
297 | |||
298 | group_cost = _cost_function(fpr, tpr, |
||
299 | base_rates[group], |
||
300 | cost_matrix) |
||
301 | |||
302 | group_cost *= proportions[group] |
||
303 | |||
304 | total_cost += group_cost |
||
305 | |||
306 | return -total_cost |
||
307 | |||
308 | acceptance_rate_min_cost = _ternary_search_float(total_cost_function, |
||
309 | 0, 1, TRINARY_SEARCH_TOL) |
||
310 | |||
311 | cost = total_cost_function(acceptance_rate_min_cost) |
||
312 | |||
313 | threshold_indices = get_acceptance_rate_indices(roc_curves, base_rates, |
||
314 | acceptance_rate_min_cost) |
||
315 | thresholds = _extract_threshold(roc_curves) |
||
316 | |||
317 | cutoffs = {group: thresholds[threshold_index] |
||
318 | for group, threshold_index |
||
319 | in threshold_indices.items()} |
||
320 | |||
321 | fpr_tpr = {group: (roc[0][threshold_indices[group]], |
||
322 | roc[1][threshold_indices[group]]) |
||
323 | for group, roc in roc_curves.items()} |
||
324 | |||
325 | return cutoffs, fpr_tpr, cost, acceptance_rate_min_cost |
||
326 | |||
327 | |||
328 | def get_fnr_indices(roc_curves, fnr_value): |
||
329 | indices = {} |
||
330 | |||
331 | tpr_value = 1 - fnr_value |
||
332 | |||
333 | for group, roc in roc_curves.items(): |
||
334 | tprs = roc[1] |
||
335 | index = _first_index_above(tprs, |
||
336 | tpr_value) - 1 |
||
337 | index = max(0, index) |
||
338 | indices[group] = index |
||
339 | |||
340 | return indices |
||
341 | |||
342 | |||
343 | View Code Duplication | def find_fnr_thresholds(roc_curves, base_rates, proportions, |
|
|
|||
344 | cost_matrix): |
||
345 | """Compute thresholds that achieve equal FNRs and minimize cost. |
||
346 | |||
347 | Also known as **equal opportunity**. |
||
348 | |||
349 | :param roc_curves: Receiver operating characteristic (ROC) |
||
350 | by attribute. |
||
351 | :type roc_curves: dict |
||
352 | :param base_rates: Base rate by attribute. |
||
353 | :type base_rates: dict |
||
354 | :param proportions: Proportion of each attribute value. |
||
355 | :type proportions: dict |
||
356 | :param cost_matrix: Cost matrix by [[tn, fp], [fn, tp]]. |
||
357 | :type cost_matrix: sequence |
||
358 | :return: Thresholds, FPR and TPR by attribute and cost value. |
||
359 | :rtype: tuple |
||
360 | |||
361 | """ |
||
362 | |||
363 | cutoffs = {} |
||
364 | |||
365 | def total_cost_function(fnr_value): |
||
366 | # todo: move demo here + multiple cost |
||
367 | indices = get_fnr_indices(roc_curves, fnr_value) |
||
368 | |||
369 | total_cost = 0 |
||
370 | |||
371 | for group, roc in roc_curves.items(): |
||
372 | index = indices[group] |
||
373 | |||
374 | fpr = roc[0][index] |
||
375 | tpr = roc[1][index] |
||
376 | |||
377 | group_cost = _cost_function(fpr, tpr, |
||
378 | base_rates[group], |
||
379 | cost_matrix) |
||
380 | group_cost *= proportions[group] |
||
381 | |||
382 | total_cost += group_cost |
||
383 | |||
384 | return -total_cost |
||
385 | |||
386 | fnr_value_min_cost = _ternary_search_float(total_cost_function, |
||
387 | 0, 1, |
||
388 | TRINARY_SEARCH_TOL) |
||
389 | |||
390 | threshold_indices = get_fnr_indices(roc_curves, fnr_value_min_cost) |
||
391 | |||
392 | cost = total_cost_function(fnr_value_min_cost) |
||
393 | |||
394 | fpr_tpr = {group: (roc[0][threshold_indices[group]], |
||
395 | roc[1][threshold_indices[group]]) |
||
396 | for group, roc in roc_curves.items()} |
||
397 | |||
398 | thresholds = _extract_threshold(roc_curves) |
||
399 | cutoffs = {group: thresholds[threshold_index] |
||
400 | for group, threshold_index |
||
401 | in threshold_indices.items()} |
||
402 | |||
403 | return cutoffs, fpr_tpr, cost, fnr_value_min_cost |
||
404 | |||
405 | |||
406 | def _find_feasible_roc(roc_curves): |
||
407 | polygons = [Delaunay(list(zip(fprs, tprs))) |
||
408 | for group, (fprs, tprs, _) in roc_curves.items()] |
||
409 | |||
410 | feasible_points = [] |
||
411 | |||
412 | for poly in polygons: |
||
413 | for p in poly.points: |
||
414 | |||
415 | if all(poly2.find_simplex(p) != -1 for poly2 in polygons): |
||
416 | feasible_points.append(p) |
||
417 | |||
418 | return np.array(feasible_points) |
||
419 | |||
420 | |||
421 | def find_separation_thresholds(roc_curves, base_rate, cost_matrix): |
||
422 | """Compute thresholds that achieve separation and minimize cost. |
||
423 | |||
424 | Also known as **equalized odds**. |
||
425 | |||
426 | :param roc_curves: Receiver operating characteristic (ROC) |
||
427 | by attribute. |
||
428 | :type roc_curves: dict |
||
429 | :param base_rate: Overall base rate. |
||
430 | :type base_rate: float |
||
431 | :param cost_matrix: Cost matrix by [[tn, fp], [fn, tp]]. |
||
432 | :type cost_matrix: sequence |
||
433 | :return: Thresholds, FPR and TPR by attribute and cost value. |
||
434 | :rtype: tuple |
||
435 | |||
436 | """ |
||
437 | |||
438 | feasible_points = _find_feasible_roc(roc_curves) |
||
439 | |||
440 | cost, (best_fpr, best_tpr) = max((_cost_function(fpr, tpr, base_rate, |
||
441 | cost_matrix), |
||
442 | (fpr, tpr)) |
||
443 | for fpr, tpr in feasible_points) |
||
444 | cost = - cost |
||
445 | |||
446 | return {}, {'': (best_fpr, best_tpr)}, cost |
||
447 | |||
448 | |||
449 | def find_thresholds(roc_curves, proportions, base_rate, |
||
450 | base_rates, cost_matrix, |
||
451 | with_single=True, with_min_cost=True, |
||
452 | with_independence=True, with_fnr=True, |
||
453 | with_separation=True): |
||
454 | """Compute thresholds that achieve various criteria and minimize cost. |
||
455 | |||
456 | :param roc_curves: Receiver operating characteristic (ROC) |
||
457 | by attribute. |
||
458 | :type roc_curves: dict |
||
459 | :param proportions: Proportion of each attribute value. |
||
460 | :type proportions: dict |
||
461 | :param base_rate: Overall base rate. |
||
462 | :type base_rate: float |
||
463 | :param base_rates: Base rate by attribute. |
||
464 | :type base_rates: dict |
||
465 | :param cost_matrix: Cost matrix by [[tn, fp], [fn, tp]]. |
||
466 | :type cost_matrix: sequence |
||
467 | |||
468 | :param with_single: Compute single threshold. |
||
469 | :type with_single: bool |
||
470 | :param with_min_cost: Compute minimum cost thresholds. |
||
471 | :type with_min_cost: bool |
||
472 | :param with_independence: Compute independence thresholds. |
||
473 | :type with_independence: bool |
||
474 | :param with_fnr: Compute FNR thresholds. |
||
475 | :type with_fnr: bool |
||
476 | :param with_separation: Compute separation thresholds. |
||
477 | :type with_separation: bool |
||
478 | |||
479 | :return: Dictionary of threshold criteria, |
||
480 | and for each criterion: |
||
481 | thresholds, FPR and TPR by attribute and cost value. |
||
482 | :rtype: dict |
||
483 | |||
484 | """ |
||
485 | |||
486 | thresholds = {} |
||
487 | |||
488 | if with_single: |
||
489 | thresholds['single'] = find_single_threshold(roc_curves, |
||
490 | base_rates, |
||
491 | proportions, |
||
492 | cost_matrix) |
||
493 | |||
494 | if with_min_cost: |
||
495 | thresholds['min_cost'] = find_min_cost_thresholds(roc_curves, |
||
496 | base_rates, |
||
497 | proportions, |
||
498 | cost_matrix) |
||
499 | |||
500 | if with_independence: |
||
501 | thresholds['independence'] = find_independence_thresholds(roc_curves, |
||
502 | base_rates, |
||
503 | proportions, |
||
504 | cost_matrix) |
||
505 | |||
506 | if with_fnr: |
||
507 | thresholds['fnr'] = find_fnr_thresholds(roc_curves, |
||
508 | base_rates, |
||
509 | proportions, |
||
510 | cost_matrix) |
||
511 | |||
512 | if with_separation: |
||
513 | thresholds['separation'] = find_separation_thresholds(roc_curves, |
||
514 | base_rate, |
||
515 | cost_matrix) |
||
516 | |||
517 | return thresholds |
||
518 | |||
519 | |||
520 | def find_thresholds_by_attr(y_true, y_score, x_sens, |
||
521 | cost_matrix, |
||
522 | with_single=True, with_min_cost=True, |
||
523 | with_independence=True, with_fnr=True, |
||
524 | with_separation=True, |
||
525 | pos_label=None, sample_weight=None, |
||
526 | drop_intermediate=False): |
||
527 | """ |
||
528 | Compute thresholds that achieve various criteria and minimize cost. |
||
529 | |||
530 | :param y_true: Binary ground truth (correct) target values. |
||
531 | :param y_score: Estimated target score as returned by a classifier. |
||
532 | :param x_sens: Sensitive attribute values corresponded to each |
||
533 | estimated target. |
||
534 | :param cost_matrix: Cost matrix by [[tn, fp], [fn, tp]]. |
||
535 | :type cost_matrix: sequence |
||
536 | |||
537 | :param pos_label: Label considered as positive and others |
||
538 | are considered negative. |
||
539 | :param sample_weight: Sample weights. |
||
540 | :param drop_intermediate: Whether to drop some suboptimal |
||
541 | thresholds which would not appear on |
||
542 | a plotted ROC curve. |
||
543 | This is useful in order to create |
||
544 | lighter ROC curves. |
||
545 | |||
546 | :param with_single: Compute single threshold. |
||
547 | :type with_single: bool |
||
548 | :param with_min_cost: Compute minimum cost thresholds. |
||
549 | :type with_min_cost: bool |
||
550 | :param with_independence: Compute independence thresholds. |
||
551 | :type with_independence: bool |
||
552 | :param with_fnr: Compute FNR thresholds. |
||
553 | :type with_fnr: bool |
||
554 | :param with_separation: Compute separation thresholds. |
||
555 | :type with_separation: bool |
||
556 | |||
557 | :return: Dictionary of threshold criteria, |
||
558 | and for each criterion: |
||
559 | thresholds, FPR and TPR by attribute and cost value. |
||
560 | :rtype: dict |
||
561 | """ |
||
562 | # pylint: disable=too-many-locals |
||
563 | |||
564 | roc_curves = roc_curve_by_attr(y_true, y_score, x_sens, |
||
565 | pos_label, sample_weight, |
||
566 | drop_intermediate) |
||
567 | |||
568 | proportions = {value: count / len(x_sens) |
||
569 | for value, count in Counter(x_sens).items()} |
||
570 | |||
571 | if pos_label is None: |
||
572 | pos_label = 1 |
||
573 | |||
574 | base_rate = np.mean(y_true == pos_label) |
||
575 | grouped = _groupby_y_x_sens(y_true, y_score, x_sens) |
||
576 | |||
577 | base_rates = {x_sens_value: np.mean(group['y_true'] == pos_label) |
||
578 | for x_sens_value, group in grouped} |
||
579 | |||
580 | thresholds_data = find_thresholds(roc_curves, |
||
581 | proportions, |
||
582 | base_rate, |
||
583 | base_rates, |
||
584 | cost_matrix, |
||
585 | with_single, with_min_cost, |
||
586 | with_independence, with_fnr, |
||
587 | with_separation) |
||
588 | |||
589 | return thresholds_data |
||
590 | |||
591 | |||
592 | def plot_roc_curves_thresholds(roc_curves, thresholds_data, |
||
593 | aucs=None, |
||
594 | title='ROC Curves by Attribute', |
||
595 | ax=None, figsize=None, |
||
596 | title_fontsize='large', |
||
597 | text_fontsize='medium'): |
||
598 | """Generate the ROC curves by attribute with thresholds. |
||
599 | |||
600 | Based on :func:`skplt.metrics.plot_roc` |
||
601 | |||
602 | :param roc_curves: Receiver operating characteristic (ROC) |
||
603 | by attribute. |
||
604 | :type roc_curves: dict |
||
605 | :param thresholds_data: Thresholds by attribute from the |
||
606 | function |
||
607 | :func:`~responsibly.interventions |
||
608 | .threshold.find_thresholds`. |
||
609 | :type thresholds_data: dict |
||
610 | :param aucs: Area Under the ROC (AUC) by attribute. |
||
611 | :type aucs: dict |
||
612 | :param str title: Title of the generated plot. |
||
613 | :param ax: The axes upon which to plot the curve. |
||
614 | If `None`, the plot is drawn on a new set of axes. |
||
615 | :param tuple figsize: Tuple denoting figure size of the plot |
||
616 | e.g. (6, 6). |
||
617 | :param title_fontsize: Matplotlib-style fontsizes. |
||
618 | Use e.g. 'small', 'medium', 'large' |
||
619 | or integer-values. |
||
620 | :param text_fontsize: Matplotlib-style fontsizes. |
||
621 | Use e.g. 'small', 'medium', 'large' |
||
622 | or integer-values. |
||
623 | :return: The axes on which the plot was drawn. |
||
624 | :rtype: :class:`matplotlib.axes.Axes` |
||
625 | |||
626 | """ |
||
627 | |||
628 | ax = plot_roc_curves(roc_curves, aucs, |
||
629 | title, ax, figsize, title_fontsize, text_fontsize) |
||
630 | |||
631 | MARKERS = ['o', '^', 'x', '+', 'p'] |
||
632 | |||
633 | for (name, data), marker in zip(thresholds_data.items(), MARKERS): |
||
634 | label = _titlify(name) |
||
635 | ax.scatter(*zip(*data[1].values()), |
||
636 | marker=marker, color='k', label=label, |
||
637 | zorder=float('inf')) |
||
638 | |||
639 | plt.legend() |
||
640 | |||
641 | return ax |
||
642 | |||
643 | |||
644 | def plot_fpt_tpr(roc_curves, |
||
645 | title='FPR-TPR Curves by Attribute', |
||
646 | ax=None, figsize=None, |
||
647 | title_fontsize='large', text_fontsize='medium'): |
||
648 | """Generate FPR and TPR curves by thresholds and by attribute. |
||
649 | |||
650 | Based on :func:`skplt.metrics.plot_roc` |
||
651 | |||
652 | :param roc_curves: Receiver operating characteristic (ROC) |
||
653 | by attribute. |
||
654 | :type roc_curves: dict |
||
655 | :param str title: Title of the generated plot. |
||
656 | :param ax: The axes upon which to plot the curve. |
||
657 | If `None`, the plot is drawn on a new set of axes. |
||
658 | :param tuple figsize: Tuple denoting figure size of the plot |
||
659 | e.g. (6, 6). |
||
660 | :param title_fontsize: Matplotlib-style fontsizes. |
||
661 | Use e.g. 'small', 'medium', 'large' |
||
662 | or integer-values. |
||
663 | :param text_fontsize: Matplotlib-style fontsizes. |
||
664 | Use e.g. 'small', 'medium', 'large' |
||
665 | or integer-values. |
||
666 | :return: The axes on which the plot was drawn. |
||
667 | :rtype: :class:`matplotlib.axes.Axes` |
||
668 | |||
669 | """ |
||
670 | |||
671 | if ax is None: |
||
672 | fig, ax = plt.subplots(1, 1, figsize=figsize) # pylint: disable=unused-variable |
||
673 | |||
674 | ax.set_title(title, fontsize=title_fontsize) |
||
675 | |||
676 | thresholds = _extract_threshold(roc_curves) |
||
677 | |||
678 | prop_cycle = plt.rcParams['axes.prop_cycle'] |
||
679 | colors = prop_cycle.by_key()['color'] |
||
680 | |||
681 | for (group, roc), color in zip(roc_curves.items(), colors): |
||
682 | plt.plot(thresholds, roc[0], '-', |
||
683 | label='{} - FPR'.format(group), color=color) |
||
684 | plt.plot(thresholds, roc[1], '--', |
||
685 | label='{} - TPR'.format(group), color=color) |
||
686 | plt.legend() |
||
687 | |||
688 | ax.set_ylim([0.0, 1.05]) |
||
689 | ax.set_xlabel('Threshold', fontsize=text_fontsize) |
||
690 | ax.set_ylabel('Probability', fontsize=text_fontsize) |
||
691 | ax.tick_params(labelsize=text_fontsize) |
||
692 | ax.legend(fontsize=text_fontsize) |
||
693 | |||
694 | return ax |
||
695 | |||
696 | |||
697 | def plot_costs(thresholds_data, |
||
698 | title='Cost by Threshold Strategy', |
||
699 | ax=None, figsize=None, |
||
700 | title_fontsize='large', text_fontsize='medium'): |
||
701 | """Plot cost by threshold definition and by attribute. |
||
702 | |||
703 | Based on :func:`skplt.metrics.plot_roc` |
||
704 | |||
705 | :param thresholds_data: Thresholds by attribute from the |
||
706 | function |
||
707 | :func:`~responsibly.interventions |
||
708 | .threshold.find_thresholds`. |
||
709 | :type thresholds_data: dict |
||
710 | :param str title: Title of the generated plot. |
||
711 | :param ax: The axes upon which to plot the curve. |
||
712 | If `None`, the plot is drawn on a new set of axes. |
||
713 | :param tuple figsize: Tuple denoting figure size of the plot |
||
714 | e.g. (6, 6). |
||
715 | :param title_fontsize: Matplotlib-style fontsizes. |
||
716 | Use e.g. 'small', 'medium', 'large' |
||
717 | or integer-values. |
||
718 | :param text_fontsize: Matplotlib-style fontsizes. |
||
719 | Use e.g. 'small', 'medium', 'large' |
||
720 | or integer-values. |
||
721 | :return: The axes on which the plot was drawn. |
||
722 | :rtype: :class:`matplotlib.axes.Axes` |
||
723 | """ |
||
724 | |||
725 | if ax is None: |
||
726 | fig, ax = plt.subplots(1, 1, figsize=figsize) # pylint: disable=unused-variable |
||
727 | |||
728 | ax.set_title(title, fontsize=title_fontsize) |
||
729 | |||
730 | costs = {_titlify(group): cost |
||
731 | for group, (_, _, cost, *_) in thresholds_data.items()} |
||
732 | |||
733 | (pd.Series(costs) |
||
734 | .sort_values(ascending=False) |
||
735 | .plot(kind='barh', ax=ax)) |
||
736 | |||
737 | ax.set_xlabel('Cost', fontsize=text_fontsize) |
||
738 | ax.set_ylabel('Threshold', fontsize=text_fontsize) |
||
739 | ax.tick_params(labelsize=text_fontsize) |
||
740 | |||
741 | return ax |
||
742 | |||
743 | |||
744 | def plot_thresholds(thresholds_data, |
||
745 | markersize=7, |
||
746 | title='Thresholds by Strategy and Attribute', |
||
747 | xlim=None, |
||
748 | ax=None, figsize=None, |
||
749 | title_fontsize='large', text_fontsize='medium'): |
||
750 | """Plot thresholds by strategy and by attribute. |
||
751 | |||
752 | Based on :func:`skplt.metrics.plot_roc` |
||
753 | |||
754 | :param thresholds_data: Thresholds by attribute from the |
||
755 | function |
||
756 | :func:`~responsibly.interventions |
||
757 | .threshold.find_thresholds`. |
||
758 | :type thresholds_data: dict |
||
759 | :param int markersize: Marker size. |
||
760 | :param str title: Title of the generated plot. |
||
761 | :param tuple xlim: Set the data limits for the x-axis. |
||
762 | :param ax: The axes upon which to plot the curve. |
||
763 | If `None`, the plot is drawn on a new set of axes. |
||
764 | :param tuple figsize: Tuple denoting figure size of the plot |
||
765 | e.g. (6, 6). |
||
766 | :param title_fontsize: Matplotlib-style fontsizes. |
||
767 | Use e.g. 'small', 'medium', 'large' |
||
768 | or integer-values. |
||
769 | :param text_fontsize: Matplotlib-style fontsizes. |
||
770 | Use e.g. 'small', 'medium', 'large' |
||
771 | or integer-values. |
||
772 | :return: The axes on which the plot was drawn. |
||
773 | :rtype: :class:`matplotlib.axes.Axes` |
||
774 | """ |
||
775 | |||
776 | if ax is None: |
||
777 | fig, ax = plt.subplots(1, 1, figsize=figsize) # pylint: disable=unused-variable |
||
778 | |||
779 | ax.set_title(title, fontsize=title_fontsize) |
||
780 | |||
781 | # TODO: refactor! |
||
782 | df = pd.DataFrame({_titlify(key): thresholds |
||
783 | for key, (thresholds, *_) in thresholds_data.items() |
||
784 | if key != 'separation'}) |
||
785 | melted_df = pd.melt(df, var_name='Strategy', value_name='Threshold') |
||
786 | melted_df['Attribute'] = list(df.index) * len(df.columns) |
||
787 | |||
788 | sns.stripplot(y='Strategy', x='Threshold', hue='Attribute', data=melted_df, |
||
789 | jitter=False, dodge=True, size=markersize, ax=ax) |
||
790 | |||
791 | minor_locator = AutoMinorLocator(2) |
||
792 | fig.gca().yaxis.set_minor_locator(minor_locator) |
||
793 | ax.grid(which='minor') |
||
794 | |||
795 | if xlim is not None: |
||
796 | ax.set_xlim(*xlim) |
||
797 | |||
798 | ax.set_xlabel('Threshold', fontsize=text_fontsize) |
||
799 | ax.set_ylabel('Strategy', fontsize=text_fontsize) |
||
800 | ax.tick_params(labelsize=text_fontsize) |
||
801 | |||
802 | return ax |
||
803 |