Control system for learning to rank fairness
A Bayesian test of demographic parity for learning to rank may be applied to determine ranking modifications. A fairness control system receiving a ranking of items may apply Bayes factors to determine a likelihood of bias for the ranking. These Bayes factors may include a factor for determining bias in each item and a factor for determining bias in the ranking of the items. An indicator of bias may be generated using the applied Bayes factors and the fairness control system may modify the ranking if the determines likelihood of bias satisfies modification criteria for the ranking.
1. A system, comprising:
at least one processor;
a memory comprising program instructions that when executed by the at least one processor cause the at least one processor to implement a control system configured to:
receive a ranking of a plurality of items;
apply a Bayes factor to the ranking of the plurality of items to determine a likelihood of bias for the ranking of the items, the Bayes factor comprising a ratio of a first metric to a second metric, the first metric comprising a determination of a likelihood of bias in the ranking of the plurality of items and the second metric comprising determination of a likelihood of absence of bias in the ranking of the plurality of items;
determine that the likelihood of bias for the ranking of the plurality of items exceeds a modification criterion;
modify, responsive to the determination, the ranking of the plurality of items according to the modification criterion; and
output the modified ranking of the plurality of items.
2. The system of claim 1 , wherein:
each item of the plurality of items comprises a feature including an amount of bias; and
the likelihood of bias for the ranking of the plurality of items is determined with respect to the feature.
3. The system of claim 2 , wherein the modifying of the ranking of the plurality of items preserves the ranking of respective items having a same value of the feature.
4. The system of claim 1 , wherein the ranking of the plurality of items is received from a ranking classifier trained using machine learning.
5. The system of claim 1 , wherein to receive the ranking of the plurality of items, the control system is configured to perform said apply, determine and modify steps to a previously received ranking of the plurality of items.
6. A method for implementing a control system, comprising:
receiving a ranking of a plurality of items;
applying a Bayes factor to the ranking of the plurality of items to determine a likelihood of bias for the ranking of the plurality of items, the Bayes factor comprising a ratio of a first metric to a second metric, the first metric comprising a determination of a likelihood of bias in the ranking of the plurality of items and the second metric comprising determination of a likelihood of absence of bias in the ranking of the plurality of items;
determining that the likelihood of bias of the plurality of items exceeds a modification criterion;
modifying, responsive to the determination, the ranking of the plurality of items according to the modification criterion; and
outputting the modified ranking of the plurality of items.
7. The method of claim 6 , wherein:
each item of the plurality of items comprises a feature including an amount of bias, wherein the feature is multi-valued; and
the likelihood of bias of the plurality of items for the ranking of the plurality of items is determined with respect to the feature.
8. The method of claim 7 , wherein the modifying of the ranking of the plurality of items preserves the ranking of respective items having a same value of the feature.
9. The method of claim 7 , wherein:
the feature is multi-valued;
the modification criterion comprises a threshold of demographic parity with respect to the feature; and
demographic parity with respect to the feature comprises ranking items of the plurality of items with a particular value of the feature proportional to a rate of occurrence of the particular value of the feature relative to all values of the feature.
10. The method of claim 6 , wherein the ranking of the plurality of items is received from a ranking classifier trained using machine learning.
11. The method of claim 6 , wherein the receiving of the ranking of the plurality of items comprises performing said applying, determining and modifying steps to a previously received ranking of the plurality of items.
12. One or more non-transitory computer-accessible storage media storing program instructions that when executed on or across one or more processors cause the one or more processors to implement a control system to enforce fairness, comprising:
receiving a ranking of a plurality of items;
applying a Bayes factor to the ranking of the plurality of items to determine a likelihood of bias for the ranking of the plurality of items, the Bayes factor comprising a ratio of a first metric to a second metric, the first metric comprising a determination of a likelihood of bias in the ranking of the plurality of items and the second metric comprising determination of a likelihood of absence of bias in the ranking of the plurality of items;
determining that the likelihood of bias of the plurality of items exceeds a modification criterion;
modifying, responsive to the determination, the ranking of the plurality of items according to the modification criterion; and
outputting the modified ranking of the plurality of items.
13. The one or more non-transitory computer-accessible storage media of claim 12 , wherein:
each item of the plurality of items comprises a feature including an amount of bias, wherein the feature is multi-valued; and
the likelihood of bias for the ranking of the plurality of items is determined with respect to the feature.
14. The one or more non-transitory computer-accessible storage media of claim 13 , wherein the modifying of the ranking of the plurality of items preserves the ranking of respective items having a same value of the feature.
15. The one or more non-transitory computer-accessible storage media of claim 13 , wherein:
the feature is multi-valued;
the modification criterion comprises a threshold of demographic parity with respect to the feature; and
demographic parity with respect to the feature comprises ranking items of the plurality of items with a particular value of the feature proportional to a rate of occurrence of the particular value of the feature relative to all values of the feature.
16. The one or more non-transitory computer-accessible storage media of claim 12 , wherein the ranking of the plurality of items is received from a ranking classifier trained using machine learning.
17. The one or more non-transitory computer-accessible storage media of claim 12 , wherein the receiving of the ranking of the plurality of items comprises performing the applying, determining and modifying steps to a previously received ranking of the plurality of items.
18. The one or more non-transitory computer-accessible storage media of claim 12 , wherein the Bayes factor comprises a ratio of a first metric to a second metric, wherein:
the first metric comprises a determination of unfairness of the ranking; and
the second metric comprises a determination of fairness of the ranking.