SLIDE 1
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology | Source for images: Google images, iStock.
Bias and Fairness in AI/ML models Swati Gupta Assistant Professor - - PowerPoint PPT Presentation
Bias and Fairness in AI/ML models Swati Gupta Assistant Professor - - PowerPoint PPT Presentation
Bias and Fairness in AI/ML models Swati Gupta Assistant Professor School of Industrial and Systems Engineering, Georgia Institute of Technology October 25, 2018 Digital Data Flows Master Class: Emerging Technologies Machine Learning
SLIDE 2
SLIDE 3
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
“Bias. When scientific or technological decisions are based
- n a narrow set of systemic, structural or social concepts and
norms, the resulting technology can privilege certain groups and harm others.” – Nature comment
What is Bias/Fairness?
3
SLIDE 4
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
“Bias. When scientific or technological decisions are based
- n a narrow set of systemic, structural or social concepts and
norms, the resulting technology can privilege certain groups and harm others.” – Nature comment
What is Bias/Fairness?
3
SLIDE 5
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
4
SLIDE 6
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
4
SLIDE 7
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
4
SLIDE 8
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
4
SLIDE 9
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
5
Outline of the talk
Bias in the data, models and variables Fairness Metrics
Statistical measures Equity measures
Trolley Problem of Choice
SLIDE 10
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
Predictive Policing
6
[Lum, Isaac, 2016] Heat map of drug arrests made “application of analytical techniques to identify likely targets for police intervention and prevent crime or solve past crimes by making statistical predictions”
SLIDE 11
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
ML finds patterns in data
7
SLIDE 12
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
ML finds patterns in data
7
SLIDE 13
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
PredPol: crime type, time, loc
8
[Kristian Lum, William Isaac, 2016] Heat map of drug arrests made
SLIDE 14
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
Not just about collection
9
We live in a biased society, so it’s inevitable that data collected about that society will be biased: inherent bias, test data, feedback, proxies.. We live in a biased society, so it’s inevitable that data collected about that society will be biased: inherent bias, test data, feedback, proxies..
SLIDE 15
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
Not just about collection
9
We live in a biased society, so it’s inevitable that data collected about that society will be biased: inherent bias, test data, feedback, proxies.. We live in a biased society, so it’s inevitable that data collected about that society will be biased: inherent bias, test data, feedback, proxies..
SLIDE 16
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
Not just about collection
9
We live in a biased society, so it’s inevitable that data collected about that society will be biased: inherent bias, test data, feedback, proxies.. “We also found that setting the gender to female resulted in getting fewer instances of an ad related to high paying jobs than setting it to male. “ We live in a biased society, so it’s inevitable that data collected about that society will be biased: inherent bias, test data, feedback, proxies..
SLIDE 17
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
Not just about collection
9
We live in a biased society, so it’s inevitable that data collected about that society will be biased: inherent bias, test data, feedback, proxies.. We do not want such biases to propagate into systems that make life-changing decisions. Proxies: predicting crime using data on arrests, Not on incidence of crime. We live in a biased society, so it’s inevitable that data collected about that society will be biased: inherent bias, test data, feedback, proxies..
SLIDE 18
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
10
Outline of the talk
Bias in the data, models and variables Fairness Metrics
Statistical measures Equity measures
Trolley Problem of Choice
SLIDE 19
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology | Source of images: Google images, iStock.
Machine Learning Pipeline
11
Machine Learning/AI Data Data driven decisions What is the effect of these decisions on human well-being?
SLIDE 20
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
Classification
12
YES NO Hired for job or not, will re-offend or not (prison), given a loan or not.
SLIDE 21
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology | ProPublica, May 2016
13
YES Hired for job or not, will re-offend or not (prison), given a loan or not.
Is it fair to achieve highest accuracy in classification?
NO
Statistical Definitions of Fairness
SLIDE 22
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology | ProPublica, May 2016
13
YES Hired for job or not, will re-offend or not (prison), given a loan or not.
Is it fair to achieve highest accuracy in classification?
NO COMPAS Risk Score: ProPublica
Statistical Definitions of Fairness
SLIDE 23
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology | ProPublica, May 2016
13
YES Hired for job or not, will re-offend or not (prison), given a loan or not.
Is it fair to achieve highest accuracy in classification?
NO COMPAS Risk Score: ProPublica
Statistical Definitions of Fairness
SLIDE 24
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology | ProPublica, May 2016
13
YES Hired for job or not, will re-offend or not (prison), given a loan or not.
Is it fair to achieve highest accuracy in classification?
NO
Statistical Definitions of Fairness
Or is it fair to balance false positives across the groups?
SLIDE 25
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology | ProPublica, May 2016
13
YES Hired for job or not, will re-offend or not (prison), given a loan or not.
Is it fair to achieve highest accuracy in classification?
NO
Statistical Definitions of Fairness
Or is it fair to balance false positives across the groups?
SLIDE 26
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology | Wikipedia
14
YES Hired for job or not, will re-offend or not (prison), given a loan or not.
Is it fair to achieve highest accuracy in classification?
NO
Statistical Definitions of Fairness
Or is it fair to balance false positives across the groups? False negatives?
SLIDE 27
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology | Wikipedia
14
Statistical Definitions of Fairness
In fact, different stakeholders might have different points of view
SLIDE 28
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
Equity Metrics of Fairness
What about general decisions: how much loan to give? where to place an emergency room? Where to schedule deliveries?
15
Is it fair to minimize total distance travelled by any group? Group A Group B distance
SLIDE 29
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
Equity Metrics of Fairness
What about general decisions: how much loan to give? where to place an emergency room? Where to schedule deliveries?
16
Group A Group B Is it fair to minimize total distance travelled by any group?
SLIDE 30
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
Equity Metrics of Fairness
What about general decisions: how much loan to give? where to place an emergency room? Where to schedule deliveries?
16
Group A Group B Is it fair to minimize total distance travelled by any group?
SLIDE 31
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
Equity Metrics of Fairness
What about general decisions: how much loan to give? where to place an emergency room? Where to schedule deliveries?
17
Group A Group B Is it fair to minimize average distance travelled by any group (per person)?
SLIDE 32
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
Equity Metrics of Fairness
What about general decisions: how much loan to give? where to place an emergency room? Where to schedule deliveries?
17
Group A Group B Is it fair to minimize average distance travelled by any group (per person)? Group by race? income? insurance?
SLIDE 33
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology | Marsh and Schilling 1993
Equity Metrics of Fairness
18
SLIDE 34
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
Bias in the data, models and variables Fairness Metrics
Statistical measures Equity measures
Trolley Problem of Choice
19
Outline of the talk
SLIDE 35
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology
Which fairness do we want?
20
?
This has to be a collective decision we need to consciously reach at, after a deeper dive into the application.
At least 50 ways to be fair
SLIDE 36
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology | Sources: shutterstock, Lum & Isaac 2016
Which fairness do we want?
21
This has to be a collective decision we need to consciously reach at, after a deeper dive into the application.
Social Scientist: Arrest data is not a good proxy for crime data
SLIDE 37
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology | Boracas and Hardt, 2017
Which fairness do we want?
22
Lawyer/Policy maker: Cannot use protected classes for making decisions.
Race (Civil Rights Act of 1964), Color (Civil Rights Act of 1964), Religion (Civil Rights Act
- f 1964), National Origin (Civil Rights Act of
1964), Citizenship (Immigration Reform and Control Act), Age (Age discrimination in Employment Act of 1967), Pregnancy (Pregnancy Discrimination Act), Familial status (Civil Rights Act of 1968), Disability status (Rehabilitation Act of 1973; Americans with Disabilities Act of 1990), Veteran Status (Vietnam Era Veterans’ Readjustment Assistance Act of 1974; Uniformed Services Employment and Reemployment Rights Act), Genetic Information (Genetic Information Nondiscrimination Act)
Disparate Treatment v/s Impact
This has to be a collective decision we need to consciously reach at, after a deeper dive into the application.
SLIDE 38
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology | Boracas and Hardt, 2017
Which fairness do we want?
22
Lawyer/Policy maker: Cannot use protected classes for making decisions.
Race (Civil Rights Act of 1964), Color (Civil Rights Act of 1964), Religion (Civil Rights Act
- f 1964), National Origin (Civil Rights Act of
1964), Citizenship (Immigration Reform and Control Act), Age (Age discrimination in Employment Act of 1967), Pregnancy (Pregnancy Discrimination Act), Familial status (Civil Rights Act of 1968), Disability status (Rehabilitation Act of 1973; Americans with Disabilities Act of 1990), Veteran Status (Vietnam Era Veterans’ Readjustment Assistance Act of 1974; Uniformed Services Employment and Reemployment Rights Act), Genetic Information (Genetic Information Nondiscrimination Act)
Disparate Treatment v/s Impact
This has to be a collective decision we need to consciously reach at, after a deeper dive into the application.
SLIDE 39
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology | Corbett-Davies et al, 2017
Which fairness do we want?
23
Algorithm designer: awareness of protected classes can fix bias
This has to be a collective decision we need to consciously reach at, after a deeper dive into the application.
SLIDE 40
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology |Kleinberg et al 2017, Chouldechova 2017
Which fairness do we want?
24
Statistician: cannot have equal false positive, negative rates & calibration simultaneously COMPAS Debate: Northpointe v/s ProPublica
This has to be a collective decision we need to consciously reach at, after a deeper dive into the application.
SLIDE 41
Bias and Fairness in AI/ML models | Swati Gupta | Ongoing work with Jalan, Ranade, Yang, Zhuang, 2018
Which fairness do we want?
25
Optimizer: can at times have approximately fair solutions for multiple metrics together
This has to be a collective decision we need to consciously reach at, after a deeper dive into the application.
Group by race? income? insurance?
SLIDE 42
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology | hscmd.org/momcare-statistics/
Which fairness do we want?
26
Economists, Behavioral scientists, Humans-in the loop, ..
This has to be a collective decision we need to consciously reach at, after a deeper dive into the application.
SLIDE 43
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology | hscmd.org/momcare-statistics/
Which fairness do we want?
26
*
This has to be a collective decision we need to consciously reach at, after a deeper dive into the application.
SLIDE 44
Bias and Fairness in AI/ML models | Swati Gupta | Georgia Institute of Technology