This document discusses various approaches for measuring and achieving fairness in machine learning models. It summarizes research on identifying discrimination from models, removing protected features, and imposing different fairness constraints. Specifically, it finds that removing a protected feature like age can decrease model performance, redundant encodings may still encode that feature, and different fairness constraints like equalized odds come at a cost to model optimization but are important to consider.
2. Protected Features
• Case: Need to avoid Age discrimination
• So remove the Age feature from the
dataset
• Then retrain the model and evaluate
performance
3. Protected Features
• We use publicly available credit card
default data. From: The comparisons of data mining techniques for the predictive
accuracy of probability of default of credit card clients. Yeh, I. C., & Lien, C. H. (2009)
• XGBoost. 5-fold Stratified Cross-
Validation.
• Including Age feature: AUC 0.78454
• Removing Age: AUC drops to 0.78326
4. Protected Features
• Business impact of removing Age feature:
• 0.17% decrease approval rate (keeping
default rate constant).
• 0.05% increase default rate (keeping approval
rate constant)
5. Protected Features
• But “Fairness through unawareness”
does not work.
• Redundant Encodings hold information
on the protected feature.
6. Protected Features
• Use the protected feature as a target.
• We try to predict it from the remaining
features.
• XGBoost. 5-Fold Quantile Stratified
Cross-Validation.
7. Protected Features
• Baseline: Mean age for everyone.
• Result: 7.54611 Mean Absolute Error.
• Improvement: XGBoost Regression
• Result: 5.99789 Mean Absolute Error
0.3 R2 Score
8. Protected Features
• Using the other features we did better
than average guessing
• 7.54611 - 5.99789 = Bayesian Fairness
Rate
• Best we can do, without removing other
non-protected features
9. Discrimination-aware
Data Mining
• First paper (2008) to look at
discrimination in ML models. By: Dino Pedreschi, Salvatore
Ruggieri, Franco Turini
• Used simple rule mining on loan data.
• See how much of performance can be
explained by discriminating feature.
12. Equality of Opportunity
in Supervised Learning
• Paper (2016) by Google Research Moritz Hardt,
Eric Price, Nathan Srebro
• Look at groups in the protected feature
• FICO loan data
14. Equality of Opportunity
in Supervised Learning
• Every profit-optimizing model has a
threshold at which a decision is made.
• Putting fairness constraints on your
model often means losing profit.
• We can study profit for a model under
different threshold constraints.
15. Equality of Opportunity
in Supervised Learning
• Max-Profit. No Fairness Constraints.
Pick different profit-maximizing threshold
for every group.
• 100% of Max-Profit.
16. Equality of Opportunity
in Supervised Learning
• Feature blind. Requires threshold to be
the same for every group.
• 99.3% of Max-Profit
17. Equality of Opportunity
in Supervised Learning
• Equal Opportunity. Picks for each
group a threshold such that the fraction
of non-defaulting group members that
qualify for loans is the same.
• 92.8% of Max-Profit.
18. Equality of Opportunity
in Supervised Learning
• Equalized odds. Requires both the
fraction of non-defaulters that qualify for
loans and the fraction of defaulters that
qualify for loans to be constant across
groups.
• 80.2% of Max-Profit.
19. Equality of Opportunity
in Supervised Learning
• Demographic parity. Picks for each
group a threshold such that the fraction
of group members that qualify for loans
is the same.
• 69.8% of Max-Profit.
20. Conclusion
• Studying fairness is new, but important
• Fairness has a measurable cost
• Ignoring the feature may not be enough
• There are different fairness constraints
• We still need the unfair feature to detect
unfairness