Fair Regression Metrics

[xixiaoyin]

Fair Regression: Quantitative Definitions and Reduction-based Algorithms

Statistical parity: f(X) is independent of the protected attribute A.

• Discrete SA
• Did not consider accuracy
• Ludicrous when there is a causal relationship between the protected attribute and the output variable

Bounded group loss (BGL):

• Discrete SA
• Have to pick loss function

[xixiaoyin]

A Convex Framework for Fair Regression (Richard et al., 2017)

• Individual Fairness

  

• Group Fairness

  

• Hybrid notions of fairness

  
• 

• Minimizing, in which f is the fairness penalty, and use l2 regularization to prevent overfitting.

• Comparing pairs of instances (one from each group) as cross pairs, essential Individual fairness, all under the notion that similar instances should be treated similarly

• Difference is that whether unfairness can be cancelled or not

• D(yi,yj) is inversely proportional to | yi − yj |

• Does not rule out perfect predictor (Like AOD and EOD)

• Whether to use sensitive attributes in learning process does not affect much

• Transfer classification to regression by Pr[yi = 1] = exp(w · xi)/(1 + exp(w · xi))

https://arxiv.org/abs/1706.02409