-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtopics20_1.html
946 lines (802 loc) · 32.2 KB
/
topics20_1.html
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
<!doctype html>
<html>
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0, maximum-scale=1.0, user-scalable=no">
<title>Fairness in Machine Learning — Part 1</title>
<link rel="stylesheet" href="dist/reset.css">
<link rel="stylesheet" href="dist/reveal.css">
<link rel="stylesheet" href="custom_themes/sussex.css" id="theme">
<!-- Theme used for syntax highlighting of code -->
<link rel="stylesheet" href="plugin/highlight/monokai.css">
<style>
.citeme {
float: right;
color: #000000;
font-size: 12pt;
font-style: italic;
margin: 0;
}
</style>
</head>
<body>
<div id="hidden" style="display:none;"><div id="static-content">
<footer>
<p>Topics in Computer Science 2020</p>
</footer>
</div> </div>
<div class="reveal"><div class="slides">
<section>
<h1>Fairness in Machine Learning</h1>
<p>by Oliver Thomas and Thomas Kehrenberg</p>
<p>{ot44,t.kehrenberg}@sussex.ac.uk - Predictive Analytics Lab (PAL), University of Sussex, UK</p>
<img src="images/sussex_logo.svg" style="max-width: 10%; border: none;"/>
</section>
<section data-markdown><textarea data-template>
## Contents
- Part 1
- Intro and notation
- What is fairness in machine learning and why should I care?
- Algorithmic fairness definitions
- Approaches to enforce algorithmic fairness
- Part 2
- Transparency in algorithmic fairness
- Causality and Counterfactual Fairness
- Toolboxes for research
</textarea></section>
<section data-markdown><textarea data-template>
## Non-Contents
- This lecture will not "solve" the problem of fairness
</textarea></section>
<section data-markdown><textarea data-template>
## Take home messages
- Fairness is _Domain Specific_.
- If you don't think about bias, it will come back to haunt you.
- Accuracy only tells part of the story.
- Fairness is an active and exciting research topic.
</textarea></section>
<section>
<h2>Machine Learning</h2>
<blockquote>...using statistical techniques to give computer systems the ability to "learn"
(e.g., progressively improve performance on a specific task)
from data, without being explicitly programmed.
</blockquote>
</section>
<section data-markdown><textarea data-template>
## Classification
- given some input $X$, predict a class label $Y \in \\{0, 1, ..., C-1\\}$
- $X$ is usually a **vector**
- often with high number of dimensions, e.g. more than 1 million for a picture
- simplest case: **binary classification**, $Y \in \\{0, 1\\}$
- for example: is there a hot dog in this picture ($Y=1$) or not ($Y=0$)?
</textarea></section>
<section data-markdown><textarea data-template>
## Classification
- we are looking to train a function $f$ that maps $X$ to $Y$
- the output is the prediction: $\hat{Y} = f(X)$
- we want $\hat{Y}$ to be as close as possible to the label $Y$
- $f$ can be implemented as
- a neural network
- an SVM
- a decision-tree model
</textarea></section>
<section data-markdown><textarea data-template>
## Training data
- training data: a set of pairs $(x, y)$
- input data $x$ with corresponding label $y$
- we are looking for model that works well on _unseen data_
- so we try to gather training data that is _similar_ to deployment setting
</textarea></section>
<section data-markdown><textarea data-template>
## Training data
- BUT this is not always possible
- if we make predictions on data that is *very different* from the training data,
the model will perform badly
- problem if the training data does not describe reality well
</textarea></section>
<section data-markdown><textarea data-template>
## Where is machine learning used?
- hiring decisions
- bail decisions
- credit approval
- insurance premiums
<aside class="notes">
Companies are already using or planning to use machine learning for these tasks.
</aside>
</textarea></section>
<section>
<p> </p>
<p> </p>
<p> </p>
But there are problems:
</section>
<section data-background-color="#fff">
<img src="images/pp_mb.png" width=72% title="Pro-Publica - Machine Bias"/>
</section>
<section data-background-color="#fff">
<img width=72% src="images/amazon.png" title="Amazon CV Screening"/>
</section>
<section data-background-color="#fff">
<img src="images/bots-at-the-gate.png" width=40% title="Bots At The Gate"/>
</section>
<section data-background-color="#fff">
<img src="images/norman.png" width=58% title="Norman"/>
</section>
<section>
<h2>Algorithmic bias</h2>
<section>
<ul>
<li>machine learning systems are making decisions that affect humans</li>
<li>these decisions should be <b>fair</b></li>
<li>by default machine learning algorithms tend to be biased in some way</li>
<ul>
<li>why?</li>
</ul>
</ul>
</section>
<section>
<p>Because the world is complicated!</p>
<p>The data is unlikely to be perfect.</p>
<p>The developer's (often arbitrary) decisions have a downstream impact</p>
<p>Business goals don't always align with fair behaviour</p>
</section>
</section>
<section data-markdown><textarea data-template>
## Algorithmic bias
(Broadly) The problem can be divided into <ins>two categories</ins>.
Both types of bias can appear together.
- bias stemming from biased training data
- bias stemming from the algorithms themselves
</textarea></section>
<section>
<p> </p>
<p> </p>
<h1>Biased training data</h1>
</section>
<section data-markdown><textarea data-template>
## Bias in, bias out
- Databases: GIGO
- ML: BIBO
- the ML algorithm just learns what is in the training data
</textarea></section>
<section data-markdown><textarea data-template>
## Examples for bias
**Task**: generate description for images
</textarea></section>
<section data-markdown><textarea data-template>
## Sampling bias ➔ tyranny of the majority
- In the <span class="highlight">imSitu</span> situation recognition <span class="highlight">dataset</span>, the activity cooking is over 33\% more likely to involve females than males in a training set, and a trained algorithm further <span class="highlight">amplifies</span> the disparity to 68\%
<span class="citeme">Zhao et al.: Men also like shopping, EMNLP 2017</span>
<img src="images/women_also_snowboard/example3.png" title="Men also like shopping"/>
</textarea></section>
<section data-markdown><textarea data-template>
## Sampling bias ➔ tyranny of the majority
- <span class="highlight">The reason is</span>: the algorithm predicts the gender from the activity and not from looking at the person
<span class="citeme">Anne Hendricks et al.: Women also snowboard, ECCV 2018</span>
<img src="images/women_also_snowboard/example2.png" title="Women also snowboard"/>
</textarea></section>
<section data-markdown><textarea data-template>
## Desired result
- what we would want is that the algorithm only looks at the person when predicting the gender
- if the gender is not recognizable from the picture, the algorithm should be "unsure"
</textarea></section>
<section data-markdown><textarea data-template>
## Biased data
- in the previous example, certain training examples were underrepresented
- sampling bias
- other case: data is simply wrong
- e.g. data was gathered by humans who just lied
</textarea></section>
<section data-markdown><textarea data-template>
## Enforcing a fair outcome
- no matter in what way the data is biased: we want to **enforce a fair outcome**
- idea: just tell the algorithm that it should treat all groups in the same way
- question: **how do we define a fair outcome?**
- really hard question
- start with the simplest definition
</textarea></section>
<section data-markdown><textarea data-template>
## Example
- task: predict whether someone should be hired ($Y=1$) or not ($Y=0$)
- two races: <span style="color: blue">blue race</span> and <span style="color: green">green race</span>
- blue: $S=0$, green: $S=1$
- in the training set: **20% of blue** applicants were hired, **80% of green** applicants were hired
</textarea></section>
<section data-markdown><textarea data-template>
## Statistical parity
$$
P(\hat{Y}=1 | S=0) = P(\hat{Y}=1 | S=1)
$$
- $\hat{Y} \in \\{0,1\\}, S \in \\{0,1\\}$
- $S$: sensitive attribute (for example gender, race)
- $\hat{Y}$: prediction
**meaning**: the probability to get a positive prediction is the same for each group
</textarea></section>
<section data-markdown><textarea data-template>
## Statistical parity
- concrete: when evaluating the algorithm on the test set, both groups ($S=0$ and $S=1$)
should have the same number of positive predictions ($\hat{Y}=1$)
- in the hiring example from before: same percentage from both races will be hired (e.g. 50%)
</textarea></section>
<section data-markdown><textarea data-template>
## Side effects of statistical parity
- enforcing statistical parity necessarily produces **lower accuracy**
- consider this: we want to enforce 50% acceptance for the blue race but the training data only has 20% accepted
- some individuals of the blue race have to be "misclassified"
($\hat{Y} = 1$ instead of $\hat{Y} = 0$)
to make the numbers work
- not surprising because we are computing accuracy against the *biased data*
</textarea></section>
<section data-markdown><textarea data-template>
## Trade-off: Fairness – Accuracy
- usually we don't want to have too bad accuracy with respect to the data
- goal: find algorithm that produces fair result at highest possible accuracy
- otherwise it's easy: a *random classifier* is very fair (but useless)
- random classifier: just predict $\hat{Y} = 1$ 50% of the time regardless of input
</textarea></section>
<section data-markdown><textarea data-template>
## Summary of Statistical Parity
- very simple criterion
- introduces Fairness – Accuracy trade-off
- more sophisticated criteria are possible (based on causal graphs)
</textarea></section>
<section>
<p> </p>
<h1>Bias introduced by the ML algorithm</h1>
</section>
<section data-markdown><textarea data-template>
## Why would an ML algorithm introduce bias?
Consider again the hiring example:
- two features: SAT score and race (blue and green) of individuals
- task: predict if they should be hired ($Y=1$) or not ($Y=0$)
- composition of the dataset: 50% blue, 50% green. 20% of blue have $Y=1$, 80% of green have $Y=1$.
</textarea></section>
<section data-markdown><textarea data-template>
## Bias from algorithm
The dataset is heavily skewed but let's ignore that for now and just try to make accurate predictions for this dataset.
</textarea></section>
<section data-markdown><textarea data-template>
## Bias from algorithm
(Reminder: 20% of blue have $Y=1$, 80% of green have $Y=1$.)
A simple way to make relatively accurate predictions:
- ignore the SAT score
- for green individuals always predict $Y=1$
<!-- base the prediction on SAT-->
- for blue individuals always predict $Y=0$
Result: up to 80% accuracy (80% in blue subgroup and 80% in green subgroup)
</textarea></section>
<section data-markdown><textarea data-template>
## Bias from algorithm
- the dataset was already skewed but the algorithm's prediction are even more "unfair"
- this is because it's easier to just base the decision on race than to figure out the effect of the SAT
- this is an extreme case but similar things can actually happen
</textarea></section>
<section data-markdown><textarea data-template>
## Bias from algorithm
**What we do not want:** the algorithm "being lazy" in a subgroup
**What we want:** the algorithm should make equally good predictions for all subgroups
Criterion that enforces this: *Equality of Opportunity*
</textarea></section>
<section data-markdown><textarea data-template>
## Equality of Opportunity
$$
P(\hat{Y}=1 | S=0, Y=1) = P(\hat{Y}=1 | S=1, Y=1)
$$
- with $Y, \hat{Y} \in \\{0,1\\}$ and $S \in \\{0,1\\}$
**meaning**: the probability of predicting 1, given that the label is 1, should be the same for all groups
**in practice**: the TPR (true positive rate) should be the same in all groups
</textarea></section>
<section data-markdown><textarea data-template>
## Equalised Odds
$$
P(\hat{Y}=y | S=0, Y=y) = P(\hat{Y}=y | S=1, Y=y)
$$
for all possible values for $y$
- stricter version of *Equality of Opportunity*
- TPR and TNR (true negative rate) must be the same in all groups
- TPR = TP / (TP + FN), TNR = TN / (TN + FP)
</textarea></section>
<section data-markdown><textarea data-template>
## Properties
- **EoO** (Equality of Opportunity) and **EO** (Equalised Odds) both assume that the training data is correct
- a perfect classifier (that always predicts the correct label) fulfills **EoO** and **EO**d
- however: a random classifier does as well
- a random classifier achieves 50% TPR (and TNR) in all groups
</textarea></section>
<section data-markdown><textarea data-template>
## Accuracy–Fairness trade-off
- achieving EoO or EO at low accuracy is easy (see random classifier)
- here, even more than for Statistical Parity, it is important to find an algorithm with good fairness and good accuracy
</textarea></section>
<!-- Border between Thomas stuff and Oliver stuff -->
<section data-markdown><textarea data-template>
## Fairness definitions
- Discrimination in the law:
- <span class="highlight">Direct discrimination</span> with respect to intent
- <span class="highlight">Indirect discrimination</span> with respect to consequences
<span class="citeme">Article 21, EU Charter of Fundamental Rights</span>
- From the legal context to algorithmic fairness, e.g.:
- Removing direct discrimination by <span class="highlight">not using group information at prediction</span>
- Removing indirect discrimination by enforcing <span class="highlight">equality on the outcomes</span> between groups
</textarea></section>
<section>
<section data-markdown><textarea data-template>
## All statistics-based Fairness Criteria
- Statistical Parity
- Equalised Odds
- Equality of Opportunity
- Calibration by Group
</textarea></section>
<section data-markdown><textarea data-template>
## Statistical Parity
#### (**independence** based notion of fairness, $S \perp \hat{Y}$)
$$
P(\hat{Y}=1 | S=0) = P(\hat{Y}=1 | S=1)
$$
$$
Y \in \\{0,1\\} \\\\
S \in \\{0,1\\}
$$
</textarea></section>
<section data-markdown><textarea data-template>
## Equalised Odds
#### (**separation** based notion of fairness, $S \perp \hat{Y} | Y$)
$$
P(\hat{Y}=1 | S=0, Y=y) = P(\hat{Y}=1 | S=1, Y=y)
$$
$$
Y \in \\{0,1\\} \\\\
S \in \\{0,1\\}
$$
</textarea></section>
<section data-markdown><textarea data-template>
## Equality of Opportunity
#### (**separation** based notion of fairness, $S \perp \hat{Y} | Y$)
$$
P(\hat{Y}=1 | S=0, Y=1) = P(\hat{Y}=1 | S=1, Y=1)
$$
$$
Y \in \\{0,1\\} \\\\
S \in \\{0,1\\}
$$
</textarea></section>
<section data-markdown><textarea data-template>
## Calibration by Group
#### (**sufficiency** based notion of fairness, $S \perp Y | \hat{Y}$)
$$
P(Y=1 | \hat{Y}=\hat{y}, S=s) = \hat{y}
$$
$$
\hat{Y} \in \[0...1\] \\\\
S \in \\{0,1\\}
$$
</textarea></section>
</section>
<section>
<h2>Which fairness criteria to choose?</h2>
<p> </p>
<p><strong>Let's have all the fairness!</strong></p>
<p>If we are fair with regards to all notions of fair, then we're fair... right?</p>
</section>
<section>
<h2>Which fairness criteria to choose?</h2>
<p>Independence based fairness (i.e. Statistical Parity)</p>
$$
\hat{Y} \perp S
$$
<p>Separation based fairness (i.e. Equalised Odds/Opportunity)</p>
$$
\hat{Y} \perp S | Y
$$
</section>
<section>
<h2>Which fairness criteria to choose?</h2>
<p>For both to hold (in binary classification), then either </p>
<ul>
<li>$S \perp Y$, our data is fair, or</li>
<li>$\hat{Y} \perp Y$, we have a random predictor.</li>
</ul>
<p>Similarly, Sufficiency cannot hold with either notion of fairness.</p>
</section>
<section>
<h1>Illustrative Example</h1>
</section>
<section>
<h2>Which fairness criteria to choose?</h2>
<p>Consider a university, and we are in charge of administration!</p>
<p>We can only accept 50% of all applicants.</p>
<p>10,000 applicants are female and 10,000 of applicants are male.</p>
<p>We have been tasked with being fair with regard to <b>gender</b>.</p>
</section>
<section>
<h2>University Admission</h2>
<p>We have an acceptance criteria that is highly predictive of success.</p>
<p>80% of those who meet the acceptance criteria will successfully graduate.</p>
<p>Only 10% of those who don't meet the acceptance criteria will successfully graduate.</p>
</section>
<section>
<h2>University Admission</h2>
<p>As we're a good university we have a lot of applications from people who don't meet the acceptance criteria.</p>
<p>60% of female applicants meet the acceptance criteria.</p>
<p>40% of male applicants meet the acceptance criteria.</p>
<p>Remember, we can only accept 50% of all applicants</p>
<h3>What should we do?</h3>
</section>
<section>
<section>
<h3>Truth Tables</h3>
<h4>Female Applicants</h4>
<table>
<thead><tr>
<th></th>
<th>Accepted</th>
<th>Not</th>
</tr></thead>
<tbody><tr>
<td>Actually Graduate</td>
<td></td>
<td></td>
</tr>
<tr>
<td>Don't Graduate</td>
<td></td>
<td></td>
</tr></tbody>
</table>
<h4>Male Applicants</h4>
<table>
<thead><tr>
<th></th>
<th>Accepted</th>
<th>Not</th>
</tr></thead>
<tbody><tr>
<td>Actually Graduate</td>
<td></td>
<td></td>
</tr>
<tr>
<td>Don't Graduate</td>
<td></td>
<td></td>
</tr></tbody>
</table>
</section>
<section>
<h3>Truth Tables</h3>
<h4>Female Applicants</h4>
<table>
<thead><tr>
<th></th>
<th>Accepted</th>
<th>Not</th>
</tr></thead>
<tbody><tr>
<td>Actually Graduate</td>
<td>$10000 \times 0.6 \times 0.8$</td>
<td>$10000 \times 0.4 \times 0.1$</td>
</tr>
<tr>
<td>Don't Graduate</td>
<td>$10000 \times 0.6 \times 0.2$</td>
<td>$10000 \times 0.4 \times 0.9$</td>
</tr></tbody>
</table>
<h4>Male Applicants</h4>
<table>
<thead><tr>
<th></th>
<th>Accepted</th>
<th>Not</th>
</tr></thead>
<tbody><tr>
<td>Actually Graduate</td>
<td>$10000 \times 0.4 \times 0.8$</td>
<td>$10000 \times 0.6 \times 0.1$</td>
</tr>
<tr>
<td>Don't Graduate</td>
<td>$10000 \times 0.4 \times 0.2$</td>
<td>$10000 \times 0.6 \times 0.9$</td>
</tr></tbody>
</table>
</section>
<section>
<h3>Truth Tables</h3>
<h4>Female Applicants</h4>
<table>
<thead><tr>
<th></th>
<th>Accepted</th>
<th>Not</th>
</tr></thead>
<tbody><tr>
<td>Actually Graduate</td>
<td>$4800$</td>
<td>$400$</td>
</tr>
<tr>
<td>Don't Graduate</td>
<td>$1200$</td>
<td>$3600$</td>
</tr></tbody>
</table>
<h4>Male Applicants</h4>
<table>
<thead><tr>
<th></th>
<th>Accepted</th>
<th>Not</th>
</tr></thead>
<tbody><tr>
<td>Actually Graduate</td>
<td>$3200$</td>
<td>$600$</td>
</tr>
<tr>
<td>Don't Graduate</td>
<td>$800$</td>
<td>$5400$</td>
</tr></tbody>
</table>
</section>
<section>
<h2>University Admission</h2>
<p>Our current system satisfies calibration-by-group!</p>
$$Y \perp S | \hat{Y}$$
</section>
</section>
<section>
<p>How would we solve this problem being fair using Statistical Parity as our measure?</p>
<section>???</section>
<section>
<p>Select 50% of applicants of both female and male applicants</p>
<p>10% of qualified female applicants are being rejected whilst an additional 10% of unqualified males are being accepted.</p>
</section>
<section>
<h4>Female Applicants</h4>
<table>
<thead><tr>
<th></th>
<th>Accepted</th>
<th>Not</th>
</tr></thead>
<tbody><tr>
<td>Actually Graduate</td>
<td>$5000 \times 0.8$</td>
<td>$(1000 \times 0.8) + (4000 \times 0.1)$</td>
</tr>
<tr>
<td>Don't Graduate</td>
<td>$5000 \times 0.2$</td>
<td>$(1000 \times 0.2) + (4000 \times 0.9)$</td>
</tr></tbody>
</table>
<h4>Male Applicants</h4>
<table>
<thead><tr>
<th></th>
<th>Accepted</th>
<th>Not</th>
</tr></thead>
<tbody><tr>
<td>Actually Graduate</td>
<td>$(4000 \times 0.8) + (1000 \times 0.1)$</td>
<td>$5000 \times 0.1$</td>
</tr>
<tr>
<td>Don't Graduate</td>
<td>$(4000 \times 0.2) + (1000 \times 0.9)$</td>
<td>$5000 \times 0.9$</td>
</tr></tbody>
</table>
</section>
<section>
<h4>Female Applicants</h4>
<table>
<thead><tr>
<th></th>
<th>Accepted</th>
<th>Not</th>
</tr></thead>
<tbody><tr>
<td>Actually Graduate</td>
<td>$4000$</td>
<td>$1200$</td>
</tr>
<tr>
<td>Don't Graduate</td>
<td>$1000$</td>
<td>$3800$</td>
</tr></tbody>
</table>
<h4>Male Applicants</h4>
<table>
<thead><tr>
<th></th>
<th>Accepted</th>
<th>Not</th>
</tr></thead>
<tbody><tr>
<td>Actually Graduate</td>
<td>$3300$</td>
<td>$500$</td>
</tr>
<tr>
<td>Don't Graduate</td>
<td>$1700$</td>
<td>$4500$</td>
</tr></tbody>
</table>
</section>
<section>
<h4>Female Applicants</h4>
<table>
<thead><tr>
<th></th>
<th>Accepted</th>
<th>Not</th>
</tr></thead>
<tbody><tr>
<td>Actually Graduate</td>
<td style="color: red">-800</td>
<td style="color: green">800</td>
</tr>
<tr>
<td>Don't Graduate</td>
<td style="color: red">-200</td>
<td style="color: green">200</td>
</tr></tbody>
</table>
<h4>Male Applicants</h4>
<table>
<thead><tr>
<th></th>
<th>Accepted</th>
<th>Not</th>
</tr></thead>
<tbody><tr>
<td>Actually Graduate</td>
<td style="color: green">100</td>
<td style="color: red">-100</td>
</tr>
<tr>
<td>Don't Graduate</td>
<td style="color: green">900</td>
<td style="color: red">-900</td>
</tr></tbody>
</table>
</section>
</section>
<section>
<p>How would we solve this problem being fair using Equal Opportunity as our measure?</p>
<section>???</section>
<section>$$TPR = \frac{TP}{TP + FN}$$</section>
<section>
<table>
<thead><tr>
<th></th>
<th>Accepted</th>
<th>Not</th>
</tr></thead>
<tbody><tr>
<td>Actually Graduate</td>
<td>$TP$</td>
<td>$FN$</td>
</tr>
<tr>
<td>Don't Graduate</td>
<td>$FP$</td>
<td>$TN$</td>
</tr></tbody>
</table>
</section>
<section>Select 55.5% of female applicants and 44.5% of male applicants, giving a TPR of 85.4% for both groups.</section>
<section>
<h4>Female Applicants</h4>
<table>
<thead><tr>
<th></th>
<th>Accepted</th>
<th>Not</th>
</tr></thead>
<tbody><tr>
<td>Actually Graduate</td>
<td>$4440$</td>
<td>$760$</td>
</tr>
<tr>
<td>Don't Graduate</td>
<td>$1110$</td>
<td>$3690$</td>
</tr></tbody>
</table>
<h4>Male Applicants</h4>
<table>
<thead><tr>
<th></th>
<th>Accepted</th>
<th>Not</th>
</tr></thead>
<tbody><tr>
<td>Actually Graduate</td>
<td>$3245$</td>
<td>$555$</td>
</tr>
<tr>
<td>Don't Graduate</td>
<td>$1205$</td>
<td>$4995$</td>
</tr></tbody>
</table>
</section>
<section>
<h4>Female Applicants</h4>
<table>
<thead><tr>
<th></th>
<th>Accepted</th>
<th>Not</th>
</tr></thead>
<tbody><tr>
<td>Actually Graduate</td>
<td style="color: red">-360</td>
<td style="color: green">360</td>
</tr>
<tr>
<td>Don't Graduate</td>
<td style="color: red">-90</td>
<td style="color: green">90</td>
</tr></tbody>
</table>
<h4>Male Applicants</h4>
<table>
<thead><tr>
<th></th>
<th>Accepted</th>
<th>Not</th>
</tr></thead>
<tbody><tr>
<td>Actually Graduate</td>
<td style="color: green">45</td>
<td style="color: red">-45</td>
</tr>
<tr>
<td>Don't Graduate</td>
<td style="color: green">405</td>
<td style="color: red">-405</td>
</tr></tbody>
</table>
</section>
</section>
<section data-markdown><textarea data-template>
## All statistics-based Fairness Criteria
- Statistical Parity
- Equalised Odds
- Equality of Opportunity
- Calibration by Group
</textarea></section>
<section data-markdown><textarea data-template>
## Contents
- Part 1
- Intro and notation
- What is fairness in machine learning and why should I care?
- Algorithmic fairness definitions
- Approaches to enforce algorithmic fairness
- Part 2
- Transparency in algorithmic fairness
- Causality and Counterfactual Fairness
- Toolboxes for research
</textarea></section>
<section data-markdown><textarea data-template>
## Take home messages
- Fairness is _Domain Specific_.
- If you don't think about bias, it will come back to haunt you.
- Accuracy only tells part of the story.
- Fairness is an active and exciting research topic.
</textarea></section>
</div></div>
<script type="module" src="settings.js"></script>
</body>
</html>