-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathpaf.html
307 lines (277 loc) · 12.9 KB
/
paf.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
<!DOCTYPE html>
<html lang="en" xmlns="http://www.w3.org/1999/html">
<title>An Algorithmic Framework for </title>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0, maximum-scale=1.0, user-scalable=no">
<script src="custom_themes/html_elements.js"></script>
<link rel="stylesheet" href="dist/reveal.css">
<link rel="stylesheet" href="custom_themes/sussex2.css" id="theme">
<div class="reveal">
<div class="slides">
<div id="background-template">
<footer>
<p>An Algorithmic Framework for Positive Action</p>
</footer>
</div>
<!-- <section class="dark-cyan">-->
<!-- <p> -->
<!-- <h2>An Algorithmic Framework<br>for Positive Action</h2>-->
<!-- <p> -->
<!-- <p style="color: white">Oliver Thomas</p>-->
<!-- <p style="color: white">Predictive Analytics Lab<br>-->
<!-- University of Sussex-->
<!-- <p> -->
<!-- <p><img src="images/logos/University_of_Sussex_Logo.svg2000_white.png" style="width: 5rem;" alt="">-->
<!-- </section>-->
<section>
<section>
<p>
<h3>An Algorithmic Framework for Positive Action</h3>
<p>
<p>Oliver Thomas<sup>1</sup>, Miri Zilka<sup>1,2</sup></p>
<p>Adrian Weller<sup>2</sup>, Novi Quadrianto<sup>1</sup></p>
<p> </p>
<p>
<small>1. Predictive Analytics Lab, University of Sussex</small><br>
<small>2. University of Cambridge</small>
</p>
<img src="images/paf/joint_logo.png" width=30% alt="">
<!-- <p> </p>-->
<flexbox style="display:flex; align-items: center; justify-content: right;padding-right: 5%">
<img src="images/logos/Twitter.png" width="30">
<small>@wearepal_ai</small>
<img src="images/logos/GitHub.png" width="30">
<small>predictive-analytics-lab</small>
</flexbox>
</section>
</section>
<!-- <section>-->
<!-- <h2>Contents</h2>-->
<!-- <p> </p>-->
<!-- <p>Positive Action</p>-->
<!-- <p>Our Approach</p>-->
<!-- <p>Alternative Approaches</p>-->
<!-- <p>Experiments</p>-->
<!-- <p>Limitations & Future Directions</p>-->
<!-- </section>-->
<!-- <section>-->
<!-- <h3>Equality Act 2010</h3>-->
<!-- <p> </p>-->
<!-- <p>UK law that protects people from discrimination in the workplace.</p>-->
<!-- <p style="text-align: left; padding-left: 9%">Protected characteristics include:</p>-->
<!-- <multi-col>-->
<!-- <one-col>-->
<!-- <ul>-->
<!-- <li>age</li>-->
<!-- <li>disability</li>-->
<!-- <li>gender reassignment</li>-->
<!-- <li>marriage and civil partnership</li>-->
<!-- <li>pregnancy and maternity</li>-->
<!-- </ul>-->
<!-- </one-col>-->
<!-- <one-col>-->
<!-- <ul>-->
<!-- <li>race</li>-->
<!-- <li>religion or belief</li>-->
<!-- <li>sex</li>-->
<!-- <li>sexual orientation</li>-->
<!-- </ul>-->
<!-- </one-col>-->
<!-- </multi-col>-->
<!-- </section>-->
<section>
<h3>Positive Action</h3>
<multi-col>
<one-col>
<h4>What</h4>
<p>Described in UK Law (Equality Act 2010).</p>
<p>Lawful measures taken to encourage and train people from under-represented groups.</p>
<!-- to help them overcome disadvantages in competing with other applicants.</p>-->
<p>Must be ‘proportionate’ to the extent and longevity of the under-representation.</p>
</one-col>
<one-col>
<h4>Why</h4>
<p>Some groups are under-represented in a particular setting.</p>
<p>Positive Action is a method to address this legally.</p>
<h4>Limitation</h4>
<p>Some positive action steps can be applied generally, for others there is a <em>cost</em> associated. </p>
</one-col>
</multi-col>
</section>
<section>
<h3>Examples</h3>
<ul>
<li>Including statements in job adverts to encourage applications from under-represented groups, such as 'we welcome female applicants'.</li>
<li><mark>Offering training or internships to help certain groups get opportunities or progress at work.</mark></li>
<li>Offering shadowing or mentoring to groups with particular needs.</li>
<li>Hosting an open day specifically for under-represented groups to encourage them to get into a particular field.</li>
<li>Favouring the job candidate from an under-represented group, where two candidates are 'as qualified as' each other.</li>
</ul>
</section>
<section>
<h4>Algorithmic Fairness Context</h4>
<multi-col>
<one-col>
<h4>Demographic Parity</h4>
<img src="images/paf/DP.png" width="80%" style="margin-left:auto;margin-right:auto;display:block">
<p style="text-align: center">Balance Positive Outcome Rates</p>
<p style="text-align: center"><span style="color: red">May violate anti-discrimination law.</span></p>
<p style="text-align: center;padding-top: 15px"><span style="color: green">Status-quo Challenged.</span></p>
</one-col>
<one-col>
<h4>Equal Opportunity</h4>
<img src="images/paf/EqOp.png" width="80%" style="padding-top:25px;margin-left:auto;margin-right:auto;display:block">
<p style="text-align: center">Balance <em>True-Positive</em> Outcome Rates</p>
<p style="text-align: center"><span style="color: green">No law violation.</span></p>
<p> </p>
<p style="text-align: center"><span style="color: red">Status-quo maintained.</span></p>
</one-col>
<one-col>
<h4>Individual Fairness</h4>
<div style="display: block;">
<img src="images/paf/IF.png" width="75%" style="margin-left:auto;margin-right:auto;display:block">
</div>
<p style="text-align: center">Similar individuals must be treated similarly.</p>
<p style="text-align: center"><span style="color: green">No law violation.</span></p>
<p> </p>
<p style="text-align: center"><span style="color: red">Status-quo maintained.</span></p>
</one-col>
</multi-col>
</section>
<section>
<h3>Why not...?</h3>
<h4>Top- Rejected?</h4>
<img src="images/paf/alternative_approach.png" width="450" height="300" alt="">
<p>Only works when the data is well modelled.</p>
</section>
<section>
<h3>Why Do We Need an Algorithmic Framework?</h3>
<img src="images/paf/problem.png" width="400">
<p>Let $S$ be a protected characteristic, $X$ some input feature and $Y$ the outcome. With arrows indicating a relationship pathway between variables.</p>
<p>We want to model the <span style="color: #20b459">green</span> pathway, remove the <span style="color: #fe211a">red</span> pathway, and try to understand the effect of the <span style="color: #24b3ef">blue</span> pathway.</p>
</section>
<!-- <section>-->
<!-- <h3>Why Do We Need an Algorithmic Framework?</h3>-->
<!-- <img src="images/paf/split_problem.png" width="400">-->
<!-- <p>Let $S$ be a protected characteristic, $X$ some input feature and $Y$ the outcome. With arrows indicating a relationship pathway between variables.</p>-->
<!-- <p>We want to model the <span style="color: #20b459">green</span> pathway, remove the <span style="color: #fe211a">red</span> pathway, and try to understand the effect of the <span style="color: #24b3ef">blue</span> pathway.</p>-->
<!-- </section>-->
<section>
<h3>Counterfactual</h3>
<flexbox style="display:flex; align-items: center; justify-content: center">
<img src="images/paf/problem.png" width=20%>
<img src="images/paf/cf.png" width=35%>
<img src="images/paf/table.png" width="40%">
</flexbox>
</section>
<!-- <section>-->
<!-- <h3>Our Approach</h3>-->
<!-- <img src="images/paf/time.png" width="80%" alt="">-->
<!-- <p>We describe the world of data using language from Friedler et al.<sup>1</sup></p>-->
<!-- <p>There may be a number of sources leading to discrepancy within each space. We are particularly interested in the transition between spaces.</p>-->
<!-- <p> </p>-->
<!-- <small><sup>1</sup> Sorelle A. Friedler, Carlos Scheidegger, and Suresh Venkatasubramanian. 2021. <br> The (Im)Possibility of Fairness: Different Value Systems Require Different Mechanisms for Fair Decision Making.</small>-->
<!-- </section>-->
<section>
<h3>Why not...?</h3>
<h4>Deferral / Rejection Modelling</h4>
<multicol>
<onecol>
<!-- asdf-->
<img src="images/paf/defer_1.png" width=20%; alt="">
</onecol>
<onecol>
<!-- asdf-->
<img src="images/paf/defer_2.png" width=20%; alt="" style="padding-right: 10px">
</onecol>
<onecol>
<img style="border-left: 3px solid black;padding-left: 20px;" src="images/paf/ours_not_defer.png" width=30%; alt="">
<p>We produce an ensemble acting on "counterfactual" samples.<br>Each model can be certain, but certain for reasons influenced by a protected characteristic — this wouldn't be caught in a deferral model.</p>
</onecol>
</multicol>
</section>
<section class="dark-cyan">
<p>
<p>
<p>
<p>
<h2>Experiments</h2>
</section>
<section>
<h3>Synthetic Data</h3>
<p>Access to Ground Truth counterfactuals and outcome beyond acceptance.</p>
<div class="r-stack">
<div>
<p> </p>
<img src="images/paf/results_1.png" width="70%">
<p><small> </small></p>
</div>
<div class="fragment current-visible">
<p>Acceptance = $P(Y=1 | S=s)$</p>
<img src="images/paf/results_2.png" width="70%">
<p><small> </small></p>
</div>
<div class="fragment current-fragment">
<p>TCP|s = $P(Y=\{1,2\} | S=s, \textrm{Grad}=1)$</p>
<img src="images/paf/results_3.png" width="70%">
<p><small>Y=2 represents the Positive Action Outcome.</small></p>
</div>
</div>
</section>
<!-- <section>-->
<!-- <h3>Real Data</h3>-->
<!-- <p>No longer have access to Ground Truth counterfactual data.</p>-->
<!-- <div class="r-stack">-->
<!-- <div>-->
<!-- <p> </p>-->
<!-- <img src="images/paf/adult_results_1.png" width="70%">-->
<!-- </div>-->
<!-- <div class="fragment current-visible">-->
<!-- <p>Additional candidates make the results <em>tend</em> towards more equal outcomes.</p>-->
<!-- <img src="images/paf/adult_results_2.png" width="70%">-->
<!-- </div>-->
<!-- </div>-->
<!-- </section>-->
<section>
<h3>Limitations</h3>
<ul>
<li>We have only considered binary protected groups.</li>
<li>Hand-determined rules may not be generally applicable.</li>
<li>Hard to evaluate.</li>
</ul>
<h3>Future Directions</h3>
<ul>
<li>Characterising uncertainty over generated counterfactuals.</li>
<li>Considering more complex scenarios.</li>
<li>Incorporate budget.</li>
</ul>
</section>
<section>
<h3>Conclusion</h3>
<p> </p>
<ol>
<li>Positive Action as an additional outcome is a method to increase equal representation over time.</li>
<li>We have presented an approach based on couterfactual modelling that tries to identify positive action candidates while mitigating directly discriminatory behaviour.</li>
</ol>
</section>
<section class="dark-cyan">
<p>
<p>
<p>
<h2>Thank you</h2>
<p>Contact us for more information.</p>
<p>Twitter: @wearepal_ai</p>
<p>Web: www.wearepal.ai</p>
<flexbox style="display:flex; align-items: center; justify-content: center">
<img src="images/logos/eu.jpeg" width="100">
<p> </p>
<img src="images/logos/erc.jpeg" width="100">
</flexbox>
</section>
<!-- <section>-->
<!-- <h3>We are currently hiring</h3>-->
<!-- <p>for PhD students in fairness, uncertainty modelling and computer vision.</p>-->
<!-- </section>-->
</div>
</div>
<script type="module" src="setup.js"></script>