fair representation learning2 hispanic javascript 1 yes no 0 1 hispanic c++ 5 yes yes 1 1 hispanic...
TRANSCRIPT
![Page 1: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/1.jpg)
Fair Representation Learning
Apr 10, 2020Dr. Wei Wei, Prof. James Landay
CS 335: Fair, Accountable, and Transparent (FAccT) Deep LearningStanford University
![Page 2: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/2.jpg)
Updated Project Policies● Maximum Number of Students For Course Projects
○ We now allow up to 3 students in a project
● Project Sharing○ Project sharing between classes can be done under the permissions from the Instructors
● Reminder: Project Proposal Deadline○ Apr 22, before class○ Less than two weeks from now
![Page 3: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/3.jpg)
Recaps From the Previous Lecture● Fairness Through Unawareness
R
H
Y S
Fair ML Model
R - Race S = SkillsY - Years of Exp O = Often Goes to Mexico Market
Indirect Discrimination
R
H
Y SO
✗ ✗
Outcomes:
![Page 4: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/4.jpg)
Limitations● Processing Sensitive Features
○ Fairness through unawareness requires sensitive features to be masked out○ Not easy to do in real life○ Referred to as individual fairness criteria
![Page 5: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/5.jpg)
Outline● Major Fairness Criteria
○ Demographic Parity○ Equality of Odds/Opportunity○ FICO Case Study
● Fair Representation Learning○ Prejudice Removing Regularizer
![Page 6: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/6.jpg)
Demographic Parity● Demographic Parity Is Applied to a Group of Samples
○ Does not require features to be masked out
● A Predictor Ŷ Satisfies Demographic Parity If ○ The probabilities of positive predictions are the same regardless of whether the group is
protected○ Protected groups are identified as A = 1
![Page 7: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/7.jpg)
Comparisons
Individual Treatment Group Treatment
Protected Features A
Non-protected Features X Protected
Features = 1Protected
Features = 0
P(Ŷ | X)Demographic Parity
P(Ŷ=1 | A =1)Demographic Parity
P(Ŷ=1 | A =0)Fairness Through Unawareness
![Page 8: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/8.jpg)
Graphical Model Explanations
R
H
Y SO
H
Y SO
H
Y SO
✗
R=1 R=0
P(H | O, Y ,S) P(H =1| R=1) P(H =1| R=0)=
Individual Treatment Group Treatment
![Page 9: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/9.jpg)
SAT Score Prediction
Feldman et al, 2015
Unconstrained P(Ŷ=1 | A =1)
UnconstrainedP(Ŷ=1 | A =0)
P(Ŷ=1 | A =0) = P(Ŷ=1 | A =0)
Ŷ
![Page 10: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/10.jpg)
Issues With Demographic Parity● Correlates Too Much With the Performance of the Predictor
predictor predictor
P(Ŷ =0 | A=0) P(Ŷ =1 | A=0)
A=0
P(Ŷ =1 | A=1) P(Ŷ =0 | A=1)
A=1match
![Page 11: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/11.jpg)
Issues With Demographic Parity● Correlates Too Much With the Performance of the Predictor
predictor predictor
P(Ŷ =0 | A=0) P(Ŷ =1 | A=0)
A=0
P(Ŷ =1 | A=1) P(Ŷ =0 | A=1)
A=1match
Y=0 Y=1 Y=1 Y=0 Y=1 Y=0Y=0 Y=1
Accepted too many who are not qualified
![Page 12: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/12.jpg)
Equality of Odds (Hardt et al, 2016)● Equal Probabilities for Both Qualified/Unqualified People Across Protected
Groups
P(Ŷ =0 | A=0) P(Ŷ =1 | A=0)
A=0
P(Ŷ =1 | A=1) P(Ŷ =0 | A=1)
A=1match
Y=0 Y=1 Y=1 Y=0 Y=1 Y=0Y=0 Y=1
![Page 13: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/13.jpg)
Equality of Opportunity (Hardt et al, 2016)● Equal Probabilities for Qualified People Across Protected Groups
P(Ŷ =0 | A=0) P(Ŷ =1 | A=0)
A=0
P(Ŷ =1 | A=1) P(Ŷ =0 | A=1)
A=1match
Y=0 Y=1 Y=1 Y=0 Y=1 Y=0Y=0 Y=1
![Page 14: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/14.jpg)
Case Study on FICO● FICO Dataset
○ 301,536 TransUnion TransRisk scores from 2003○ Scores ranges from 300 to 850○ People were labeled as in default if they failed to pay a debt for at least 90 days○ Protected attribute A is race, with four values: {Asian, white non-Hispanic, Hispanic, and black}
![Page 15: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/15.jpg)
FICO Scores● 18% Default Rate on Any Accounts Corresponds to a 2% Default Rate for New Loans
![Page 16: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/16.jpg)
Making Lending Decisions Without Discriminating● Requirement: Default Rate < 18%, Simple Threshold Model
○ Max Profit - No Fairness Constraints○ Race Blind - Using the same threshold for all race groups
![Page 17: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/17.jpg)
Making Lending Decisions Without Discriminating● Requirement: Default Rate < 18%, Simple Threshold Model
○ Max Profit - No Fairness Constraints○ Race Blind - Using the same threshold for all race groups○ Demographic Parity
■ Fraction of the group members that qualify for the loan are the same
![Page 18: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/18.jpg)
Making Lending Decisions Without Discriminating● Requirement: Default Rate < 18%, Simple Threshold Model
○ Max Profit - No Fairness Constraints○ Race Blind - Using the same threshold for all race groups○ Demographic Parity
■ Fraction of the group members that qualify for the loan are the same
○ Equal Opportunity■ Fraction of non-defaulting group members that qualify for the loan is the same
![Page 19: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/19.jpg)
Making Lending Decisions Without Discriminating● Requirement: Default Rate < 18%, Simple Threshold Model
○ Max Profit - No Fairness Constraints○ Race Blind - Using the same threshold for all race groups○ Demographic Parity
■ Fraction of the group members that qualify for the loan are the same
○ Equal Opportunity■ Fraction of non-defaulting group members that qualify for the loan is the same
○ Equal Odds■ Fraction of both non-defaulting and defaulting groups of members that quality for the
loan is the same
![Page 20: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/20.jpg)
Credit Modeling Using A Single Threshold● Within-Group Percentile Differs Dramatically for Each Group
620
82%
![Page 21: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/21.jpg)
Found Thresholds for Each Fairness Definitions
![Page 22: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/22.jpg)
Identifying Non-Defaulters
![Page 23: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/23.jpg)
Non-Defaulters and Max Profits
![Page 24: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/24.jpg)
Practice Question● Find out the Fairness Criteria that Ŷ1, and Ŷ2 Satisfy
○ A = {race}, Y = {Hiring Decision}
Race and Ethnicity
Skill Years of Exp
Goes to Mexican Markets?
Hiring Decision Y
PredictorŶ1
PredictorŶ2
Hispanic Javascript 1 yes no 0 1
Hispanic C++ 5 yes yes 1 1
Hispanic Python 1 no yes 1 0
White Java 2 no yes 0 0
White C++ 3 no yes 1 1
White C++ 0 no no 1 0
![Page 25: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/25.jpg)
● P(Ŷ1 = 1 | R = H) = ● P(Ŷ1 = 1 | R = W) =
Demographic Parity for Predictor Ŷ1
Race and Ethnicity
Skill Years of Exp
Goes to Mexican Markets?
Hiring Decision Y
PredictorŶ1
PredictorŶ2
Hispanic Javascript 1 yes no 0 1
Hispanic C++ 5 yes yes 1 1
Hispanic Python 1 no yes 1 0
White Java 2 no yes 0 0
White C++ 3 no yes 1 1
White C++ 0 no no 1 0
2/3
![Page 26: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/26.jpg)
● P(Ŷ1 = 1 | R = H) = ● P(Ŷ1 = 1 | R = W) =
Demographic Parity for Predictor Ŷ1
Race and Ethnicity
Skill Years of Exp
Goes to Mexican Markets?
Hiring Decision Y
PredictorŶ1
PredictorŶ2
Hispanic Javascript 1 yes no 0 1
Hispanic C++ 5 yes yes 1 1
Hispanic Python 1 no yes 1 0
White Java 2 no yes 0 0
White C++ 3 no yes 1 1
White C++ 0 no no 1 0
2/32/3
![Page 27: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/27.jpg)
● P(Ŷ1 = 1 | R = H) = ● P(Ŷ1 = 1 | R = W) =
Demographic Parity for Predictor Ŷ1
Race and Ethnicity
Skill Years of Exp
Goes to Mexican Markets?
Hiring Decision Y
PredictorŶ1
PredictorŶ2
Hispanic Javascript 1 yes no 0 1
Hispanic C++ 5 yes yes 1 1
Hispanic Python 1 no yes 1 0
White Java 2 no yes 0 0
White C++ 3 no yes 1 1
White C++ 0 no no 1 0
2/32/3
Demographics Parity✔
![Page 28: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/28.jpg)
● P(Ŷ1 = 1 | R = H, Y = yes) = ● P(Ŷ1 = 1 | R = W, Y = yes) = ● P(Ŷ1 = 1 | R = H, Y = no) = ● P(Ŷ1 = 1 | R = W, Y = no) =
Equality of Opportunity/Odds for Predictor Ŷ1
Race and Ethnicity
Skill Years of Exp
Goes to Mexican Markets?
Hiring Decision Y
PredictorŶ1
PredictorŶ2
Hispanic Javascript 1 yes no 0 1
Hispanic C++ 5 yes yes 1 1
Hispanic Python 1 no yes 1 0
White Java 2 no yes 0 0
White C++ 3 no yes 1 1
White C++ 0 no no 1 0
1
![Page 29: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/29.jpg)
● P(Ŷ1 = 1 | R = H, Y = yes) = ● P(Ŷ1 = 1 | R = W, Y = yes) = ● P(Ŷ1 = 1 | R = H, Y = no) = ● P(Ŷ1 = 1 | R = W, Y = no) =
Equality of Opportunity/Odds for Predictor Ŷ1
Race and Ethnicity
Skill Years of Exp
Goes to Mexican Markets?
Hiring Decision Y
PredictorŶ1
PredictorŶ2
Hispanic Javascript 1 yes no 0 1
Hispanic C++ 5 yes yes 1 1
Hispanic Python 1 no yes 1 0
White Java 2 no yes 0 0
White C++ 3 no yes 1 1
White C++ 0 no no 1 0
10.5
![Page 30: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/30.jpg)
● P(Ŷ1 = 1 | R = H, Y = yes) = ● P(Ŷ1 = 1 | R = W, Y = yes) = ● P(Ŷ1 = 1 | R = H, Y = no) = ● P(Ŷ1 = 1 | R = W, Y = no) =
Equality of Opportunity/Odds for Predictor Ŷ1
Race and Ethnicity
Skill Years of Exp
Goes to Mexican Markets?
Hiring Decision Y
PredictorŶ1
PredictorŶ2
Hispanic Javascript 1 yes no 0 1
Hispanic C++ 5 yes yes 1 1
Hispanic Python 1 no yes 1 0
White Java 2 no yes 0 0
White C++ 3 no yes 1 1
White C++ 0 no no 1 0
10.5
0
![Page 31: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/31.jpg)
● P(Ŷ1 = 1 | R = H, Y = yes) = ● P(Ŷ1 = 1 | R = W, Y = yes) = ● P(Ŷ1 = 1 | R = H, Y = no) = ● P(Ŷ1 = 1 | R = W, Y = no) =
Equality of Opportunity/Odds for Predictor Ŷ1
Race and Ethnicity
Skill Years of Exp
Goes to Mexican Markets?
Hiring Decision Y
PredictorŶ1
PredictorŶ2
Hispanic Javascript 1 yes no 0 1
Hispanic C++ 5 yes yes 1 1
Hispanic Python 1 no yes 1 0
White Java 2 no yes 0 0
White C++ 3 no yes 1 1
White C++ 0 no no 1 0
10.5
01
![Page 32: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/32.jpg)
● P(Ŷ1 = 1 | R = H, Y = yes) = ● P(Ŷ1 = 1 | R = W, Y = yes) = ● P(Ŷ1 = 1 | R = H, Y = no) = ● P(Ŷ1 = 1 | R = W, Y = no) =
Equality of Opportunity/Odds for Predictor Ŷ1
Race and Ethnicity
Skill Years of Exp
Goes to Mexican Markets?
Hiring Decision Y
PredictorŶ1
PredictorŶ2
Hispanic Javascript 1 yes no 0 1
Hispanic C++ 5 yes yes 1 1
Hispanic Python 1 no yes 1 0
White Java 2 no yes 0 0
White C++ 3 no yes 1 1
White C++ 0 no no 1 0
10.5
01
Equality of Opportunity
Equality of Odds
✗
✗
![Page 33: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/33.jpg)
● P(Ŷ1 = 1 | R = H) = ● P(Ŷ1 = 1 | R = W) =
Demographic Parity for Predictor Ŷ2
Race and Ethnicity
Skill Years of Exp
Goes to Mexican Markets?
Hiring Decision Y
PredictorŶ1
PredictorŶ2
Hispanic Javascript 1 yes no 0 1
Hispanic C++ 5 yes yes 1 1
Hispanic Python 1 no yes 1 0
White Java 2 no yes 0 0
White C++ 3 no yes 1 1
White C++ 0 no no 1 0
2/3
![Page 34: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/34.jpg)
● P(Ŷ1 = 1 | R = H) = ● P(Ŷ1 = 1 | R = W) =
Demographic Parity for Predictor Ŷ2
Race and Ethnicity
Skill Years of Exp
Goes to Mexican Markets?
Hiring Decision Y
PredictorŶ1
PredictorŶ2
Hispanic Javascript 1 yes no 0 1
Hispanic C++ 5 yes yes 1 1
Hispanic Python 1 no yes 1 0
White Java 2 no yes 0 0
White C++ 3 no yes 1 1
White C++ 0 no no 1 0
2/31/3
![Page 35: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/35.jpg)
● P(Ŷ1 = 1 | R = H) = ● P(Ŷ1 = 1 | R = W) =
Demographic Parity for Predictor Ŷ2
Race and Ethnicity
Skill Years of Exp
Goes to Mexican Markets?
Hiring Decision Y
PredictorŶ1
PredictorŶ2
Hispanic Javascript 1 yes no 0 1
Hispanic C++ 5 yes yes 1 1
Hispanic Python 1 no yes 1 0
White Java 2 no yes 0 0
White C++ 3 no yes 1 1
White C++ 0 no no 1 0
2/31/3
Demographics Parity✗
![Page 36: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/36.jpg)
● P(Ŷ1 = 1 | R = H, Y = yes) = ● P(Ŷ1 = 1 | R = W, Y = yes) = ● P(Ŷ1 = 1 | R = H, Y = no) = ● P(Ŷ1 = 1 | R = W, Y = no) =
Equality of Opportunity/Odds for Predictor Ŷ2
Race and Ethnicity
Skill Years of Exp
Goes to Mexican Markets?
Hiring Decision Y
PredictorŶ1
PredictorŶ2
Hispanic Javascript 1 yes no 0 1
Hispanic C++ 5 yes yes 1 1
Hispanic Python 1 no yes 1 0
White Java 2 no yes 0 0
White C++ 3 no yes 1 1
White C++ 0 no no 1 0
1/2
![Page 37: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/37.jpg)
● P(Ŷ1 = 1 | R = H, Y = yes) = ● P(Ŷ1 = 1 | R = W, Y = yes) = ● P(Ŷ1 = 1 | R = H, Y = no) = ● P(Ŷ1 = 1 | R = W, Y = no) =
Equality of Opportunity/Odds for Predictor Ŷ2
Race and Ethnicity
Skill Years of Exp
Goes to Mexican Markets?
Hiring Decision Y
PredictorŶ1
PredictorŶ2
Hispanic Javascript 1 yes no 0 1
Hispanic C++ 5 yes yes 1 1
Hispanic Python 1 no yes 1 0
White Java 2 no yes 0 0
White C++ 3 no yes 1 1
White C++ 0 no no 1 0
1/21/2
![Page 38: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/38.jpg)
● P(Ŷ1 = 1 | R = H, Y = yes) = ● P(Ŷ1 = 1 | R = W, Y = yes) = ● P(Ŷ1 = 1 | R = H, Y = no) = ● P(Ŷ1 = 1 | R = W, Y = no) =
Equality of Opportunity/Odds for Predictor Ŷ2
Race and Ethnicity
Skill Years of Exp
Goes to Mexican Markets?
Hiring Decision Y
PredictorŶ1
PredictorŶ2
Hispanic Javascript 1 yes no 0 1
Hispanic C++ 5 yes yes 1 1
Hispanic Python 1 no yes 1 0
White Java 2 no yes 0 0
White C++ 3 no yes 1 1
White C++ 0 no no 1 0
1/21/2
1
![Page 39: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/39.jpg)
● P(Ŷ1 = 1 | R = H, Y = yes) = ● P(Ŷ1 = 1 | R = W, Y = yes) = ● P(Ŷ1 = 1 | R = H, Y = no) = ● P(Ŷ1 = 1 | R = W, Y = no) =
Equality of Opportunity/Odds for Predictor Ŷ2
Race and Ethnicity
Skill Years of Exp
Goes to Mexican Markets?
Hiring Decision Y
PredictorŶ1
PredictorŶ2
Hispanic Javascript 1 yes no 0 1
Hispanic C++ 5 yes yes 1 1
Hispanic Python 1 no yes 1 0
White Java 2 no yes 0 0
White C++ 3 no yes 1 1
White C++ 0 no no 1 0
1/21/2
10
![Page 40: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/40.jpg)
● P(Ŷ1 = 1 | R = H, Y = yes) = ● P(Ŷ1 = 1 | R = W, Y = yes) = ● P(Ŷ1 = 1 | R = H, Y = no) = ● P(Ŷ1 = 1 | R = W, Y = no) =
Equality of Opportunity/Odds for Predictor Ŷ2
Race and Ethnicity
Skill Years of Exp
Goes to Mexican Markets?
Hiring Decision Y
PredictorŶ1
PredictorŶ2
Hispanic Javascript 1 yes no 0 1
Hispanic C++ 5 yes yes 1 1
Hispanic Python 1 no yes 1 0
White Java 2 no yes 0 0
White C++ 3 no yes 1 1
White C++ 0 no no 1 0
1/21/2
10
Equality of Opportunity
Equality of Odds
✔
✗
✔
✗
![Page 41: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/41.jpg)
Summary of Fairness Criteria
Fairness Criteria Criteria Group Individual
Unawareness Excludes A in Predictions ✓
Demographic Parity
Equalized Odds ✓
Equalized Opportunity ✓
![Page 42: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/42.jpg)
Outline● Major Fairness Criteria
○ Demographic Parity○ Equality of Odds/Opportunity○ FICO Case Study
● Fair Representation Learning○ Prejudice Removing Regularizer
![Page 43: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/43.jpg)
Fair Representation Learning● Make Representations Fair
○ Ensure fairness up to a certain level
x
z
y
fair
![Page 44: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/44.jpg)
Prejudice Remover Regularizer (Kamishima et al, 2012)
● Quantified Causes of Unfairness○ Prejudice
■ Unfairness rooted in the dataset○ Underestimation
■ Model unfairness because the model is not fully converged○ Negative Legacy
■ Unfairness due to sampling biases
● Training Objective
Loss of the Model Fairness Regularizer L2 Regularizer
![Page 45: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/45.jpg)
Prejudice Index (PI)● Recall that Indirect Discrimination Happens When
○ Prediction is not directly conditioned on sensitive variables S○ Prediction is indirectly conditioned on S by a variable O that is dependent on S○ P(Ŷ | O), and O ~ P(O | S)
● Prejudice Index (PI)○ Measures the degree of indirect discrimination based on mutual information
little prejudicesome prejudiceKamishima et al, 2012prediction model
![Page 46: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/46.jpg)
Normalized Prejudice Index (NPI)● Prejudice Index (PI)
○ Measures the degree of indirect discrimination based on mutual information○ Ranges in [0, +∞)
● Normalized Prejudice Index (NPI)○ Normalize PI by the entropy of Y and S○ Ranges in [0, 1]
Kamishima et al, 2012
![Page 47: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/47.jpg)
Optimizing PI ● Learning PI
Kamishima et al, 2012
![Page 48: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/48.jpg)
Optimizing PI ● Learning PI
● Using Logistic Regression Model as the Prediction Model
Prediction Model
Kamishima et al, 2012
double summations triple summations
Expands Pr(Y, S) into ΣxPr(X, Y, S)
![Page 49: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/49.jpg)
Optimizing PI ● Learning PI
● Using Logistic Regression Model as the Prediction Model
Kamishima et al, 2012
difficult to evaluate
![Page 50: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/50.jpg)
Optimizing PI
Kamishima et al, 2012
![Page 51: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/51.jpg)
Optimizing PI
Kamishima et al, 2012
Integrals Are Difficult to Evaluate
![Page 52: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/52.jpg)
Optimizing PI
Kamishima et al, 2012
Approximating integrals by sample means
![Page 53: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/53.jpg)
Optimizing PI
Kamishima et al, 2012
Approximating integrals by sample means
![Page 54: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/54.jpg)
Putting Things Together● Optimization Target
● Fairness Regularizer
Loss of the Model Fairness Regularizer L2 Regularizer
Kamishima et al, 2012
![Page 57: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/57.jpg)
Results● Changes of Performance With η
○ Model performance decreases (Acc)○ Discrimination Decreases (NPI)○ "Fairness Efficiency" (PI/MI) Increases
Kamishima et al, 2012
![Page 58: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/58.jpg)
Adult Income Dataset (Kohavi 1996)● Predict Whether Income Exceeds $50K/yr Based on Census Data
![Page 60: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/60.jpg)
Results● Prejudice Prior Sacrifices Model Performance
○ PR has lower Acc (Accuracy) ○ PR has lower NMI (normalized mutual information between labels and predictions)
● Prejudice Prior Makes Model Fair○ PR has lower NPI
Kamishima et al, 2012η is the weight we put on prejudice regularizers
Logistic Regressionfull fet.Logistic Regressionno sensitive fet.
Logistic Regression + Prejudice Regularizer
![Page 61: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/61.jpg)
Results● PI/MI
○ Prejudice Index / Mutual Information ○ Demonstrates a trade-offs between model fairness and performance○ Measures the amount of discrimination we eliminate with one unit of performance gain
(measured by MI)
Kamishima et al, 2012η - weight put on the prejudice regularizer
Logistic Regressionfull fet.Logistic Regressionno sensitive fet.
Logistic Regression + Prejudice Regularizer
![Page 62: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/62.jpg)
Reading Assignments (Pick One)● A. Beutel, J. Chen, Z. Zhao, and E. H. Chi, Data decisions and theoretical implications when
adversarially learning fair representations, FAT 2017● Kleinberg, Jon, Sendhil Mullainathan, and Manish Raghavan. Inherent trade-offs in the fair
determination of risk scores, ArXiv, 2016● Depeng Xu, Shuhan Yuan, Lu Zhang, and Xintao Wu. Fairgan: Fairness-aware generative
adversarial networks. IEEE International Conference on Big Data (Big Data), 2018 ● Creager, E., Madras, D., Jacobsen, J. H., Weis, M. A., Swersky, K., Pitassi, T., & Zemel, R.
Flexibly fair representation learning by disentanglement, ICML 2019● Jiang, R., Pacchiano, A., Stepleton, T., Jiang, H., & Chiappa, S. Wasserstein Fair
Classification. UAI, 2019
![Page 63: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/63.jpg)
Next Lecture
Interpretability and Transparency
![Page 64: Fair Representation Learning2 Hispanic Javascript 1 yes no 0 1 Hispanic C++ 5 yes yes 1 1 Hispanic Python 1 no yes 1 0 White Java 2 no yes 0 0 White C++ 3 no yes 1 1 White C++ 0 no](https://reader035.vdocument.in/reader035/viewer/2022071018/5fd2503f6271396ef017bf59/html5/thumbnails/64.jpg)
Questions?