Path7.Mod1.e - Responsible AI Dashboard - UnFairness Mitigation Algorithms

Question 1

R PP

The Two Fairlearn Unfairness Mitigation Algorithm Types; describe each

Answer 1

• Reduction: Take a standard black-box ML estimator and generate a set of retrained models using a sequence of reweighted training datasets (ex. up or downweight gender). Pick the model that gives the best trade-off w.r.t. fairness vs accuracy
• Post-Processing: Take an existing classifier and a Sensetive Feature as input. Derive a transformation of the classifier’s prediction to enforce the specified Fairness Constraints

Question 2

EG GSCl GSRe TO

The four Fairlearn Unfairness Mitigation Algorithms

Answer 2

• ExponentiatedGradient
• GridSearch (Classification Variant)
• GridSearch (Regression Variant)
• ThreshholdOptimizer

Question 3

Which Unfairness Mitigation Algorithms are Reduction/Black Box Approaches vs Post-Processing Algorithms

Answer 3

Only ThreshholdOptimizer is a Post-Processing Algorithm. The rest are Reduction Algorithms

Equality of Opportunity in Supervised Learning

Question 4

Which Unfairness Mitigation Algorithms are suited for Binary Classification vs Regression

Answer 4

Only the GridSearch Regression variant is suited for Regression tasks. The rest are suited for Binary Classification

Question 5

Which Unfairness Mitigation Algorithms support Demographic Parity and Equalized Odds Parity Constraints vs Bounded Group Loss Parity Constraints

Answer 5

Only the GridSearch Regression variant supports the Bounded Group Loss Parity Constraint. The rest support Demographic Parity and Equalized Odds

Question 6

Which Unfairness Mitigation Algorithms support Binary Sensitive vs Categorical Sensitive Features

Answer 6

Both GridSearch variants are suited for Binary Sensetive Features. Otherwise ExponentiatedGradient and ThresholdOptimizer are suited for Categorical Sensetive Features

Question 7

Which Unfairness Mitigation Algorithms hold the advantage of simplicity and flexibility as it doesn’t require retraining the model.

Answer 7

ThreshholdOptimizer doesn’t require retraining the model, instead derives a transformation…

Remember…Reduction Algorithms retrain models