fairness in ml 2: equal opportunity and odds€¦ · • when positive outcome (1) is desirable •...
TRANSCRIPT
![Page 1: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/1.jpg)
Fairness in ML 2:Equal opportunity and odds
Privacy & Fairness in Data ScienceCompSci 590.01 Fall 2018
Slides adapted from https://fairmlclass.github.io/4.html
![Page 2: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/2.jpg)
Outline
• Observational measure of fairness– Issues with Disparate Impact– Equal opportunity and Equalized odds– Positive Rate Parity– Tradeoff
• Achieving Equalized Odds– Binary Classifier
2
![Page 3: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/3.jpg)
Supervised Learning3
X1 … … … … Race Bail
0 … 0 1 … 1 Y
1 … 1 0 … 1 N
1 … 1 0 … 0 N
.. … … … … … …
X (features)Y (label)
A (protected attribute)
![Page 4: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/4.jpg)
Demographic parity(or the reverse of disparate impact)
4
![Page 5: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/5.jpg)
Demographic parity Issues5
A = 1
A = 0
Y = 1
![Page 6: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/6.jpg)
Demographic parity Issues
• Does not seem “fair” to allow random performance on A = 0
• Perfect classification is impossible
6
A = 1
A = 0
✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔
✔ ✔ ✔ ✔
![Page 7: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/7.jpg)
Perfect Classifier and Fairness
• The perfect classifier may not ensure demographic parity– Y is correlated with A
• What if we did not know how the classifier C was created?– No access to the classifier (to retrain) – No access to the training data (human created
classifier)
7
![Page 8: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/8.jpg)
True Positive Parity (TPP) (or equal opportunity)
• When positive outcome (1) is desirable• Equivalently, primary harm is due to false
negatives– Deny bail when person will not recidivate
8
![Page 9: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/9.jpg)
TPP
• Forces similar performance on Y = 1
9
A = 1
A = 0
✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔
✔✔
![Page 10: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/10.jpg)
False Positive Parity (FPP)
• TPP + FPP: Equalized Odds, orPositive Rate Parity
R satisfies equalized odds if R is conditionally independent of A given Y.
10
![Page 11: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/11.jpg)
Positive Rate Parity11
A = 1
A = 0
✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔
✔
✔
✔ ✔
![Page 12: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/12.jpg)
Predictive Value Parity
Equalized chance of success given acceptance
12
![Page 13: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/13.jpg)
Predictive Value Parity13
A = 1
A = 0
✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔
✔
✔
✔ ✔
𝑃" 𝑌 = 1 𝐶 = 1] =
𝑃) 𝑌 = 1 𝐶 = 1] =
𝑃" 𝑌 = 1 𝐶 = 0] =
𝑃) 𝑌 = 1 𝐶 = 0] =
![Page 14: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/14.jpg)
Predictive Value Parity14
A = 1
A = 0
✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔
✔
✔
✔ ✔
𝑃" 𝑌 = 1 𝐶 = 1] = 8/9
𝑃) 𝑌 = 1 𝐶 = 1] =1/3
𝑃" 𝑌 = 1 𝐶 = 0] = 0
𝑃) 𝑌 = 1 𝐶 = 0] = 0
![Page 15: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/15.jpg)
Trade-off
• We will look at a similar result later in the course due to Kleinberg, Mullainathan and Raghavan (2016)
15
![Page 16: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/16.jpg)
Outline
• Observational measure of fairness– Issues with Disparate Impact– Equal opportunity and Equalized odds– Positive Rate Parity– Tradeoff
• Achieving Equalized Odds– Binary Classifier
16
![Page 17: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/17.jpg)
Equalized Odds
R satisfies equalized odds if R is conditionally independent of A given Y.
• Derived Classifier: A new classifier 𝐶. that only depends on C, A (and Y)
17
![Page 18: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/18.jpg)
Derived Classifier18
A = 1
A = 0
✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔
✔
✔
✔
𝑃" 𝐶 = 1 𝑌 = 0] ≠ 𝑃) 𝐶 = 1 𝑌 = 0]
![Page 19: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/19.jpg)
Derived Classifier• Options for 𝐶.:– 𝐶. = 𝐶– 𝐶. = 1 − 𝐶– 𝐶. = 1– 𝐶. = 0– Or some randomized
combination of these
19
𝑃" 𝐶 = 1 𝑌 = 0]𝑃 "𝐶=1𝑌=1]
0.0 1.00.5
0.0
1.0
0.5
+
-
1
o
𝐶. is in the enclosed region
![Page 20: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/20.jpg)
Derived Classifier20
𝐶. is in this region for A = 0
𝐶. is in this region for A = 1
𝑃1 𝐶 = 1 𝑌 = 0]
𝑃 1𝐶=1𝑌=1]
![Page 21: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/21.jpg)
Summary: Multiple fairness measures
• Demographic parity or disparate impact– Pro: Used in the law– Con: Perfect classification is impossible– Achieved by modifying training data
• Equal Odds/ Opportunity– Pro: Perfect classification is possible– Con: Different groups can get rates of positive
prediction– Achieved by post processing the classifier
21
![Page 22: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/22.jpg)
Summary: Multiple fairness measures
• Equal odds/opportunity– Different groups may be treated unequally– Maybe due to the problem– Maybe due to bias in the dataset
• While demographic parity seems like a good fairness goal for the society, …Equal odds/opportunity seems to be measuring whether an algorithm is fair (independent of other factors like input data).
22
![Page 23: Fairness in ML 2: Equal opportunity and odds€¦ · • When positive outcome (1) is desirable • Equivalently, primary harm is due to false negatives – Deny bail when person](https://reader036.vdocuments.site/reader036/viewer/2022071212/60250c44a0b24c40e725ff2b/html5/thumbnails/23.jpg)
Summary: Multiple fairness measures
• Fairness through Awareness:– Need to define a distance function d(x,x’)– A guarantee at the individual level (rather than on
groups)– How does this connect to other notions of fairness?
23