-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtopics20_2.html
730 lines (606 loc) · 27.7 KB
/
topics20_2.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
<!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 data-markdown><textarea data-template>
## Individual Fairness
"Similar individual should be treated similarly"
<span class="citeme">Dwork et al. - Fairness Through Awareness</span>
<br/>
What is similar?
What is similarly?
</textarea></section>
<section>
<h2>Fairness based on similarity</h2>
<ul>
<li>First define a distance metric on your datapoints (i.e. how similar are the datapoints)
<ul>
<li>Can be just Euclidean distance but is usually something else (because of different scales)</li>
<li>This is the hardest step and requires domain knowledge</li>
</ul>
</li>
</ul>
</section>
<section>
<h2>Pre-processing based on similarity</h2>
<ul>
<li>An individual is then considered to be unfairly treated if it is treated differently than its "neighbours".</li>
<li>For any data point we can check how many of the $k$ nearest neighbours have the same class label as that data point</li>
<li>If the percentage is under a certain threshold then there was discrimination against the individual corresponding to that data point.</li>
<li>Then: flip the class labels of those data points where the class label is considered unfair</li>
</ul>
</section>
<section>
<h2>Considering similarity during training</h2>
<p>Alternative idea:</p>
<ul>
<li>a classifier is fair if and only if the predictive distributions for any two data points are at least as similar
as the two points themselves
<ul>
<li>(according to a given similarity measure for distributions and a given similarity measure for data points)</li>
</ul>
</li>
</ul>
</section>
<section>
<h2>Considering similarity during training</h2>
<ul>
<li>Needed similarity measure for distributions</li>
<li>Then add fairness condition as additional loss term to optimization loss</li>
</ul>
</section>
<section>
<h2>Which fairness criteria should we use?</h2>
<p>We've seen several statistical definitions and Individual Fairness</p>
<p>
There's no right answer, all of the previous examples are "fair".
It's important to consult <span class="highlight">domain experts</span> to find which is the best fit for each problem.
</p>
<p>
There is no one-size fits all.
</p>
</section>
<section>
<h2>Delayed Impact of Fair Learning</h2>
<p>In the real world there are implications.</p>
<p>An individual doesn't just cease to exist after we've made our loan or bail decision.</p>
<p>The decision we make has <b>consequences</b>.</p>
</section>
<section data-markdown><textarea data-template>
## Delayed Impact of Fair Learning
Consider a bank's lending decision.
The outcome is not simply reject or accept the applicant for a loan. In fact, there are multiple effects of this decision.
If the applicant receives a loan, then goes on to successfully pay it back, then not only will the bank make a profit, but then the applicant's credit score will increase.
This will make future loan decisions more favorable to the applicant.
</textarea></section>
<section data-markdown><textarea data-template>
## Delayed Impact of Fair Learning
We could think of this in terms of loss functions.
The banks objective was to maximize profit. After deciding to give the loan, the loan was repaid (with interest). The bank made money.
To maximize the objective (minimize the negative of the loss) the bank wants to give out as many loans that are likely to be repaid as possible.
</textarea></section>
<section data-markdown><textarea data-template>
## Delayed Impact of Fair Learning
In this scenario we have a slightly different concern. We don't want loans to be given to one subgroup (such as blue race), more often than members of another subgroup (green race).
Our concern is that if the number of people who receive loans from one group outweigh the other, then as a whole, the credit rating will become a euphemism for race.
</textarea></section>
<section>
<section>
<h2>The Outcome Curve</h2>
<img src="images/outcome_curve.png" width="65%">
</section>
<section>
<h2>Possible areas</h2>
<table>
<tr>
<th>Area</th>
<th>Description</th>
</tr>
<tr>
<td>Active Harm</td>
<td>Expected change in credit score of an individual is negative</td>
</tr>
<tr>
<td>Stagnation</td>
<td>Expected change in credit score of an individual is 0</td>
</tr>
<tr>
<td>Improvement</td>
<td>Expected change in credit score of an individual is positive</td>
</tr>
</table>
</section>
<section>
<h2>Possible areas</h2>
<table>
<tr>
<th>Area</th>
<th>Description</th>
</tr>
<tr>
<td>Relative Harm</td>
<td>Expected change in credit score of an individual is less than if the selection policy had been to maximize profit</td>
</tr>
<tr>
<td>Relative Improvement</td>
<td>Expected change in credit score of an individual is better than if the selection policy had been to maximize profit</td>
</tr>
</table>
</section>
</section>
<section data-markdown><textarea data-template>
## How to enforce fairness?
Fairness constraints can be added <span class="highlight">pre</span>, <span class="highlight">during</span> or <span class="highlight">post</span> training.
<span class="highlight">Pre-training</span> examples
- Zemel Fair Representations
- Beutel Adversarial Representation
- Quadrianto Fair Interpretable Representations
- Kehrenberg Null-sampling
</textarea></section>
<section data-markdown><textarea data-template>
## How to enforce fairness?
<span class="highlight">During</span> Training
- Zafar's methods
- Zhang Gradient Projection
- Kehrenberg Fair GP
<span class="highlight">Post</span> Training
- Hardt Recalibration
</textarea></section>
<section data-markdown><textarea data-template>
More on those shortly...
</textarea></section>
<section data-markdown><textarea data-template>
# Simple Approaches
</textarea></section>
<section>
<section>
<h2>Removing a sensitive feature</h2>
<p>"If we just remove the sensitive feature, then our model can't be unfair"</p>
<p></p>
<p>This doesn't work, why?</p>
</section>
<section>
<h2>Removing a sensitive feature</h2>
<p>Because ML methods are excellent at finding <span class="highlight">patterns in data</span></p>
<p><span class="citeme">Pedreschi et al.</span></p>
</section>
</section>
<section>
<section>
<h2>Reweighing</h2>
<p>Kamiran & Calders determined that one source of unfairness was imbalanced training data.</p>
<p>Simply count the current distribution of demographics</p>
<img src="images/reweighing_before.png"/>
</section>
<section>
<h2>Reweighing</h2>
<p>Then either up/down-sample or assign instance weights to members of each group in the training set
so that the results are "normalised".</p>
<img src="images/reweighing_after.png"/>
</section>
</section>
<section data-markdown><textarea data-template>
## Fair Representations
Popular approach is to add produce a "fair" representation. Consider that we have 2 roles, a data vendor, who is charge of collecting the data and preparing it. Our other role is a data user, someone who will be making predictions based on our data.
The data vendor is concerned that the data user may be using their data to make unfair decisions. So the data vendor decides to learn a new, fair representation.
</textarea></section>
<section data-markdown><textarea data-template>
The state-of-the-art method a fair representation this is to use an *adversarial* network, using a "Gradient-Reversal Layer" from Ganin et al.
<span class="citeme">Ganin et al. 2016</span>
</textarea></section>
<section>
<section>
<h2>Problems with doing this?</h2>
<h4>Any Ideas?</h4>
</section>
<section>
<h2>Problems with doing this?</h2>
<h4>What does this representation mean?</h4>
<p>The learned representation is uninterpretable by default. Recently Quadrianto et al constrained the representation to be in the same same as the input so that we could look at what changed</p>
</section>
<section>
<h2>Problems with doing this?</h2>
<h4>What if the vendor data user decides to be fair as well?</h4>
<p>Referred to as "fair pipelines". Work has only just begun exploring these. Current research shows that these don't work (at the moment!)</p>
</section>
</section>
<section>
<section>
<h2>Variational Fair Autoencoder</h2>
<p>idea: Let's "disentangle" the sensitive attribute using the variational autoencoder framework!</p>
</section>
<section>
<h2>Variational Fair Autoencoder</h2>
<img src="images/vfae_graphical_model.png">
<p><span class="citeme">Louizos et al. 2017</span> </p>
</section>
<section>
<h2>Variational Fair Autoencoder</h2>
<p>Where $z_1$ and $z_2$ are encouraged to confirm to a prior distribution</p>
</section>
<section>
<h2>Problems with doing this?</h2>
<p>Similar to adversarial model, but more principled.</p>
</section>
</section>
<section>
<h2>How to enforce fairness?</h2>
<h3>During Training</h3>
Instead of building a fair representation, we just make the fairness constraints part of the objective during training of the model.
An early example of this is by Zafar et al.
</section>
<section>
<h2>How to enforce fairness?</h2>
<h3>During Training</h3>
<p>Given we have a loss function, $\mathcal{L}(\theta)$.</p>
<p>In an unconstrained classifier, we would expect to see</p>
$$
\min{\mathcal{L}(\theta)}
$$
</section>
<section>
<h2>How to enforce fairness?</h2>
<h3>During Training</h3>
<!-- <section> -->
<p>To reduce Disparate Impact, Zafar adds a constraint to the loss function.</p>
$$
\begin{aligned}
\text{min } & \mathcal{L}(\theta) \\
\text{subject to } & P(\hat{y} \neq y|s = 0) − P(\hat{y} \neq y|s = 1) \leq \epsilon \\
\text{subject to } & P(\hat{y} \neq y|s = 0) − P(\hat{y} \neq y|s = 1) \geq -\epsilon
\end{aligned}
$$
<!-- </section> -->
<!-- <section>
<p>Where $\{{s_i \} }_{i=1}^{N}$ denotes a user's sensitive attributes</p>
<p>$\{ {d_{\theta}(x_i) \} }_{i=1}^{N}$ the signed distance to the decision boundary</p>
<p>$c$ is the trade-off between accuracy and fairness.</p>
<p>If $c$ is sufficiently small, this is the equivalent of</p>
$$
P(d_{\theta}(x) \geq 0 | s = 0) = P(d_{\theta}(x) \geq 0 | s = 1)
$$
</section>
<section>
In other words, the decisions for both sensitive groups must be equally distributed across the decision boundary up to some threshold $c$.
</section> -->
</section>
<section>
<h1>Transparency and Fairness</h1>
</section>
<section data-markdown><textarea data-template>
## Pre-processing
- Many <span class="highlight">adversarial-based</span> fair representation learning approaches, e.g. using a "Gradient-Reversal Layer"
<img src="images/models/adversarialfair.png" width=38% title="Adversarially Fair"/>
<span class="citeme">Ganin et al.,Domain-adversarial training of neural networks, JMLR 2016</span>
<div style="height:20px;font-size:1px;"> </div>
<span class="citeme">Edwards and Storkey, Censoring representations with an adversary, ICLR 2016</span>
<div style="height:20px;font-size:1px;"> </div>
<span class="citeme">Beutel et al., Data decisions and theoretical implications when adversarially learning fair representations, Jul 2017</span>
<span class="citeme">Madras et al., Learning adversarially fair and transferable representations, ICML 2018</span>
</textarea></section>
<section data-markdown><textarea data-template>
## Problems?
Can we visualise/understand the latent representation $z$?
</textarea></section>
<section>
<section data-markdown><textarea data-template>
## Contrastive examples
- The <span class="highlight">ideal dataset</span> contains an imaginary data point for each person, i.e. the one inside the black box, whereby we <span class="highlight">intervene</span> and set the gender attribute to the opposite that is in real life
<span class="citeme">Thomas, NQ, Imagined Examples for Fairness: Sampling Bias and Proxy Labels, Sep 2019</span>
<span class="citeme">Sharmanska, Hendricks, Darrell, NQ, Contrastive examples for addressing the tyranny of the majority, May 2019</span>
<img src="images/balanced_Page_1.png" width=50% title="Balanced1"/>
<img src="images/balanced_Page_2.png" width=30% title="Balanced2"/>
</textarea></section>
<section data-markdown><textarea data-template>
## Contrastive on Adult Income dataset
- Change in the <span class="highlight">relationship</span> and <span class="highlight">marital status</span>, i.e. from Husband and Married-civ-spause in real samples to Unmarried and/or Wife in contrastive samples
<img src="images/adult_real.png" width=30% title="Real"/>
<img src="images/adult_contrastive.png" width=30% title="Contrastive"/>
<img src="images/adult_diff.png" width=30.5% title="Difference"/>
</textarea></section>
<section data-markdown><textarea data-template>
## Contrastive on Adult Income dataset
- Connections to <span class="highlight">counterfactuals</span> based on the Nabi \& Shpitser's (2018) causal graph
- To remove bias towards females, the direct effect of gender on income, as well as, the <span class="highlight">effect of gender on income through marital status have to be suppressed</span>
- Experimental results
<small>
<table>
<tbody><tr>
<td></td>
<td width="10%">Accuracy $\uparrow$</td>
<td width="10%">TPR Diff. $\downarrow$</td>
<td width="10%">FPR Diff. $\downarrow$</td>
</tr>
<tr>
<td width="10%">LR (Real)</td>
<td width="10%">$85.16\pm0.14$ </td>
<td width="10%">$7.98\pm1.52$ </td>
<td width="10%">$7.23\pm0.41$</td>
</tr>
<tr>
<td width="10%">Kamiran & Calders LR (Real)</td>
<td width="10%">$84.37\pm0.28$ </td>
<td width="10%">$14.3\pm1.16$ </td>
<td width="10%">$1.17\pm0.29$ </td>
</tr>
<tr>
<td width="20%">LR (Real and NN contrastive)</td>
<td width="10%">$85.01\pm0.25$ </td>
<td width="10%">$14.80\pm1.90$ </td>
<td width="10%">$8.20\pm0.51$ </td>
</tr>
<tr>
<td width="20%">LR (Real and GAN contrastive)</td>
<td width="10%">$82.48\pm0.44$ </td>
<td width="10%">$4.95\pm3.67$ </td>
<td width="10%">$3.94\pm1.33$ </td>
</tr>
</tbody>
</table>
</small>
</textarea></section>
<section data-markdown><textarea data-template>
## Interpretability with contrastive examples
- All previous work with adversarial learning try to <span class="highlight">remove</span> sensitive attributes from data
- Instead, we use adversarial learning to <span class="highlight">generate</span> data points with pre-specified sensitive attributes (contrastive examples)
- Contrastive examples can be easily interpreted because they do have the semantic meaning of the input
</textarea></section>
</section>
<section>
<h2>Fairness without constraints</h2>
<ul>
<li>Most fairness methods for "during training" are formulated as constraints
<li>This means either
<ul>
<li>Using very special models where constrained-based optimization is possible
<li>Using Lagrange multipliers
</ul>
</ul>
</section>
<section>
<h2>Lagrange multipliers</h2>
<p>Constrained optimization problem:
<p>$$\min\limits_\theta\, f(\theta) \quad\text{s.th.}~ g(\theta) = 0$$
<p>This gets reformulated as
<p>$f(\theta) + \lambda g(\theta)$
<p>where $\lambda$ has to be determined, but is usually just set to some value
</section>
<section>
<h2>Shortcomings of this</h2>
<ul>
<li>The value of $\lambda$ has no real meaning and is usually just set to arbitrary values
<li>It's hard to tell whether the constraint is fulfilled
</ul>
</section>
<section>
<h2>Alternate approach</h2>
<ul>
<li>Introduce <em>pseudo labels</em> $\bar{y}$.
<li>Make sure $\bar{y}$ is fair ($\bar{y}\perp s$).
<li>Use $\bar{y}$ as prediction target.
</ul>
<p>$$P(y=1|x,s,\theta)=\sum\limits_{\bar{y}\in\{0,1\}} P(y=1|\bar{y},s)P(\bar{y}|x,\theta)$$
</section>
<section>
<h2>Alternate approach</h2>
<p>For demographic parity:
<p>$P(\bar{y}|s=0)=P(\bar{y}|s=1)$
<p>This allows us to compute $P(y=1|\bar{y},s)$. Which is all we need.
<p>Details: <em>Tuning Fairness by Balancing Target Labels</em> (Kehrenberg et al., 2020)
</section>
<section>
<h2>Multinomial sensitive attribute</h2>
<p>The usual metrics for Demographic Parity are only defined for binary attributes: it's either a <em>ratio</em> or a <em>difference</em> of two probabilities.
<p><b>Alternative</b>:
<p>Hirschfeld-Gebelein-Rényi Maximum Correlation Coefficient (HGR)
<ul>
<li>defined on the domain [0,1]
<li>$HGR(Y,S) = 0$ iff $Y\perp S$
<li>$HGR(Y,S) = 1$ iff there is a deterministic function to map between them
</ul>
</section>
<section>
<h2>Causality</h2>
<p><b>Motivation</b>:
<ul>
<li>Fairness methods so far have only looked at <em>correlations</em> between the sensitive attribute and the prediction
<li>But isn't a causal path more important?
</ul>
</section>
<section>
<h2>Causality</h2>
<p>Example:
<p>Admissions at Berkeley college.
<p>Men were found to be accepted with much higher probability. Was it discrimination?
</section>
<section>
<h2>Causality</h2>
<p>Was it discrimination?
<p>Not necessarily! The reason was that women were applying to more competitive departments.
<p>Departments like medicine and law are very competitive (hard to get in).
<p>Departments like computer science are much less competetive (because it's boring ;).
</section>
<section>
<h2>Causality</h2>
<ul>
<li>$R$: admission decision
<li>$A$: gender
<li>$X$: department choice
</ul>
<p>
<img src="images/berkeley_admission.png" style="width: 30%">
</section>
<section>
<h2>Causality</h2>
<!-- <p>"Correlation doesn't imply causation"</p> -->
<!-- <p class="fragment">But what is causation?</p> -->
<p>If we can understand what <b>causes</b> unfair behavior, then we can take steps to mitigate it.</p>
<p class="fragment">Basic idea: a sensitive attribute may only affect the prediction via legitimate causal paths.</p>
<p class="fragment">But how do we model causation?</p>
</section>
<section>
<h2>Causal Graphs</h2>
<p><b>Solution:</b> build causal graphs of your problem</p>
<p><b>Problem:</b> causality cannot be inferred from observational data</p>
<ul>
<li>Observational data can only show correlations
<li>For causal information we have to do <i>experiments</i>. (But that is often not ethical.)
<li>Currently: just guess the causal structure
</ul>
</section>
<section>
<h2>Path-specific causal fairness</h2>
<p>
$A$ is a sensitive attribute, and its direct effect on $Y$ and effect through $M$ is considered unfair.
But $L$ is considered admissible.
<p><img src="images/path-specific-fairness.png" style="width: 70%">
</section>
<section>
<h2>Path-specific causal fairness</h2>
<p>With a structural causal model (SCM), like this:
<p><img src="images/scm.png" style="width: 50%">
<p>You can figure out exactly how each feature should be incorporated in order to make a fair prediction.
</section>
<section>
<h2>Path-specific causal fairness</h2>
<ul>
<li><b>Advantage:</b> we can figure out <em>exactly</em> how to avoid unfairness
<li><b>Problem:</b> requires an SCM, which we usually don't have and which is hard to get
<li><b>Problem:</b> we have to decide for each feature individually whether it's admissible
<ul>
<li>This will not work for computer vision tasks where the raw features are pixels
</ul>
</ul>
</section>
<section>
<h2>Counterfactual Fairness</h2>
<ul>
<li>slightly different idea but also based on SCMs (structural causal model)
<li>we're basically asking: in the counterfactual world where your gender was different, would you have been accepted?
<ul>
<li><em>counterfactual</em>: something that has not happened or is not the case
</ul>
</ul>
</section>
<section>
<h2>Counterfactuals</h2>
<p>Example of a counterfactual statement:
<blockquote>If Oswald had not shot Kennedy, no-one else would have.</blockquote>
<p>This is counterfactual, because Oswald did in fact shoot Kennedy.
<p>It's a claim about a counterfactual world.
</section>
<section>
<h2>Counterfactual Fairness</h2>
<p>$U$: set of all unobserved background variables</p>
<p>$P(\hat{y}_{s=i}(U) = 1|x, s=i)=P(\hat{y}_{s=j}(U) = 1|x, s=i)$</p>
<p>$i, j \in \{0, 1\}$</p>
<p>$\hat{y}_{s=k}$: prediction in the counterfactual world where $s=k$</p>
<p><b>practical consequence:</b> $\hat{y}$ is counterfactually fair if it does not causally depend
(as defined by the causal model) on $s$ or any descendants of $s$.</p>
</section>
<section>
<h2>The problem of getting the SCM</h2>
<ul>
<li>Causal models cannot be learned from observational data only
<ul>
<li>You can try, but there is no <em>unique</em> solution
</ul>
<li>Different causal models can lead to very different results
</ul>
</section>
<section>
<h2>Example with Causal Graphs</h2>
<p><b>Example:</b> Law school success</p>
<p>Task: given GPA score and LSAT (law school entry exam), predict grades after one year in law school:
FYA (first year average)</p>
<p>Additionally two sensitive attributes: <i>race</i> and <i>gender</i></p>
</section>
<section>
<h2>Example with Causal Graphs</h2>
<p>Two possible graphs</p>
<img src="./images/lsat_causal.png"/>
</section>
<section>
<h2>Causality — Summary</h2>
<ul>
<li>Causal models are very promising and arguably the "right" way to solve the problem.
<li>However these models are difficult to obtain.
</ul>
</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>