Optimized score transformation for fair classification
Obtain, from an existing machine learning classifier, original probabilistic scores classifying samples taken from two or more groups into two or more classes via supervised machine learning. Associate the original probabilistic scores with a plurality of original Lagrange multipliers. Adjust values of the plurality of original Lagrange multipliers via low-dimensional convex optimization to obtain updated Lagrange multipliers that satisfy fairness constraints as compared to the original Lagrange multipliers. Based on the updated Lagrange multipliers, closed-form transform the original probabilistic scores into transformed probabilistic scores that satisfy the fairness constraints while minimizing loss in utility. The fairness constraints are with respect to the two or more groups.
1. A method comprising:
obtaining, from an existing machine learning classifier, original probabilistic scores classifying samples taken from two or more groups into two or more classes via supervised machine learning;
associating said original probabilistic scores with a plurality of original Lagrange multipliers;
adjusting values of said plurality of original Lagrange multipliers via low-dimensional convex optimization to obtain updated Lagrange multipliers that satisfy fairness constraints as compared to said original Lagrange multipliers;
based on said updated Lagrange multipliers, closed-form transforming said original probabilistic scores into transformed Lagrange multiplier-based scores that satisfy said fairness constraints while minimizing loss in utility, said fairness constraints being with respect to said two or more groups;
generating modified training data from said transformed Lagrange multiplier-based scores; and
training a new classifier on said modified training data using a standard, non-fairness-constrained algorithm;
wherein said samples comprise first samples, further comprising, with said new trained classifier, classifying second samples into said two or more classes via supervised machine learning with said new classifier.
2. The method of claim 1 , further comprising using said transformed Lagrange multiplier-based scores as classification output.
3. The method of claim 1 , further comprising:
obtaining original probabilistic scores classifying said second samples taken from said two or more groups into said two or more classes via said supervised machine learning; and
based on said updated Lagrange multipliers, closed-form transforming said original probabilistic scores classifying said second samples into transformed Lagrange multiplier-based scores that satisfy said fairness constraints while minimizing loss in utility, said fairness constraints being with respect to said two or more groups.
4. The method of claim 1 , wherein said method is carried out without knowledge of protected attributes.
5. The method of claim 1 , wherein, in said closed-form transforming step, said fairness constraints comprise mean score parity.
6. The method of claim 1 , wherein, in said closed-form transforming step, said fairness constraints comprise false positive parity.
7. The method of claim 1 , wherein, in said closed-form transforming step, said fairness constraints comprise false negative parity.
8. The method of claim 1 , wherein, in said closed-form transforming step, minimizing said loss in utility comprises minimizing cross-entropy between said transformed Lagrange multiplier-based scores and said original probabilistic scores.
9. The method of claim 1 , wherein said low-dimensional convex optimization and said closed-form transforming reduce run time in a computer carrying out score transformation.
10. The method of claim 1 , further comprising controlling distribution of at least one of electrical power, water, and computing resources in accordance with said transformed Lagrange multiplier-based scores that satisfy said fairness constraints while minimizing said loss in utility.
11. A system comprising:
a memory;
a non-transitory computer readable medium comprising computer executable instructions; and
at least one processor, coupled to said memory and said non-transitory computer readable medium, and operative to execute said instructions to be operative to:
obtain, from an existing machine learning classifier, original probabilistic scores classifying samples taken from two or more groups into two or more classes via supervised machine learning;
associate said original probabilistic scores with a plurality of original Lagrange multipliers;
adjust values of said plurality of original Lagrange multipliers via low-dimensional convex optimization to obtain updated Lagrange multipliers that satisfy fairness constraints as compared to said original Lagrange multipliers;
based on said updated Lagrange multipliers, closed-form transform said original probabilistic scores into transformed Lagrange multiplier-based scores that satisfy said fairness constraints while minimizing loss in utility, said fairness constraints being with respect to said two or more groups;
generate modified training data from said transformed Lagrange multiplier-based scores; and
train a new classifier on said modified training data using a standard, non-fairness-constrained algorithm.
12. The system of claim 11 , wherein said at least one processor is further operative to execute said instructions to use said transformed Lagrange multiplier-based scores as classification output.
13. The system of claim 11 , wherein said at least one processor is further operative to:
obtain original probabilistic scores classifying said second samples taken from said two or more groups into said two or more classes via said supervised machine learning; and
based on said updated Lagrange multipliers, closed-form transform said original probabilistic scores classifying said second samples into transformed Lagrange multiplier-based scores that satisfy said fairness constraints while minimizing loss in utility, said fairness constraints being with respect to said two or more groups.
14. The system of claim 11 , wherein said at least one processor is operative to carry out said instructions without knowledge of protected attributes.
15. The system of claim 11 , wherein minimizing said loss in utility comprises minimizing cross-entropy between said transformed Lagrange multiplier-based scores and said original probabilistic scores.
16. The system of claim 11 , wherein said at least one processor is further operative to control distribution of at least one of electrical power, water, and computing resources in accordance with said transformed Lagrange multiplier-based scores that satisfy said fairness constraints while minimizing said loss in utility.
17. A non-transitory computer readable medium comprising computer executable instructions which when executed by a computer cause the computer to perform the method of:
obtaining, from an existing machine learning classifier, original probabilistic scores classifying samples taken from two or more groups into two or more classes via supervised machine learning;
associating said original probabilistic scores with a plurality of original Lagrange multipliers;
adjusting values of said plurality of original Lagrange multipliers via low-dimensional convex optimization to obtain updated Lagrange multipliers that satisfy fairness constraints as compared to said original Lagrange multipliers;
based on said updated Lagrange multipliers, closed-form transforming said original probabilistic scores into transformed Lagrange multiplier-based scores that satisfy said fairness constraints while minimizing loss in utility, said fairness constraints being with respect to said two or more groups;
generating modified training data from said transformed Lagrange multiplier-based scores; and
training a new classifier on said modified training data using a standard, non-fairness-constrained algorithm.
18. The non-transitory computer readable medium of claim 17 , wherein the computer-readable instructions further cause the computer to perform the additional method step of controlling distribution of at least one of electrical power, water, and computing resources in accordance with said transformed Lagrange multiplier-based scores that satisfy said fairness constraints while minimizing said loss in utility.