Online post-processing in rankings for constrained utility maximization
Online post-processing may be performed for rankings generated with constrained utility maximization. A stream of data items may be received. A batch of data items from the stream may be ranked according to a ranking model trained to rank data items in a descending order of relevance. The batch of data items may be associated with a current time step. A re-ranking model may be applied to generate a re-ranking of the batch of data items according to a re-ranking policy that considers the current batch and previous batches with regard to a ranking constraint. The re-ranked items may then be sent to an application.
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 ranking system providing fairness across a plurality of rankings, the ranking system configured to:
receive a stream of data items over time;
provide, as part of continuous monitoring of the ranking system, respective fair rankings of respective pluralities of data items for a plurality of time steps including a current time step and one or more time steps earlier than the current time step, wherein to provide a fair ranking for the current time step the ranking system is configured to:
apply a ranking model to generate a ranking of a plurality of data items in a batch of data items obtained from the stream of data items, wherein the batch of data items are associated with the current time step, and wherein the ranking model is a machine learning model trained to generate the ranking of the plurality of data items in the batch in a descending order of relevance for an application;
apply a re-ranking model to generate a fair ranking of the plurality of data items in the batch, wherein an order of the fair ranking is different from the descending order of relevance for the application; wherein the re-ranking model is applied according to a re-ranking policy that satisfies a fairness constraint applicable, in aggregate, to the generated ranking for the batch of data items and respective generated fair rankings for one or more other batches of data items from the stream of data items associated with different respective time steps earlier than the current time step, wherein the one or more other batches of data items were previously ranked using the ranking model and re-ranked, using the re-ranking model, with respective fair rankings of one or more further batches with respective time steps earlier than the different respective time steps of the one or more other batches before being sent to the application, wherein the fairness constraint is determined based on parity between different groups associated with individual ones of the plurality of data items; and
send the fair ranking of the plurality of data items to the application.
2. The system of claim 1 , wherein the re-ranking policy is trained by a learning to search technique that iteratively selects data items from different queues corresponding to the different groups associated with the individual ones of the plurality of data items according to a reference policy when re-ranking one or more other batches of data items from a test data set and wherein respective deviations are made from a reference policy when making individual selections from the different queues.
3. The system of claim 1 , wherein the re-ranking policy is trained by a learning to search technique that iteratively selects data items from different queues corresponding to the different groups associated with the individual ones of the plurality of data items according to a reference policy when re-ranking one or more other batches of data items from a test data set and wherein respective deviations are made from a combination of a reference policy and a previously learned policy when making individual selections from the different queues.
4. The system of claim 1 , wherein the re-ranking model applies a deterministic re-ranking policy.
5. The system of claim 4 , wherein the deterministic re-ranking policy finds a most highly ranked data item for a protected member which is below a data item for a non-protected member in the ranking of the batch of items and swaps the data item for the protected member with the most highly ranked data item for the protected member with the data item for the non-protected member in the ranking of the batch of items.
6. A method of providing fairness across a plurality of rankings, comprising:
receiving, by a ranking system, a stream of data items over time;
providing, as part of continuous monitoring of the ranking system, respective fair rankings of respective pluralities of data items for a plurality of time steps including a current time step and one or more time steps earlier than the current time step, wherein providing a fair ranking for the current time step comprises:
applying, by the ranking system, a ranking model to generate a ranking of a plurality of data items in a batch of data items obtained from the stream of data items, wherein the batch of data items are associated with the current time step, and wherein the ranking model is a machine learning model trained to generate the ranking of the plurality of data items in the batch in a descending order of relevance for an application;
applying, by the ranking system, a re-ranking model to generate a fair ranking of the plurality of data items in the batch, wherein an order of the fair ranking is different from the descending order of relevance for the application; wherein the re-ranking model is applied according to a re-ranking policy that satisfies a ranking constraint applicable, in aggregate, to the generated ranking for the batch of data items and respective generated fair rankings for one or more other batches of data items from the stream of data items associated with different respective time steps earlier than the current time step, wherein the one or more other batches of data items were previously ranked using the ranking model and re-ranked, using the re-ranking model, with respective fair rankings of one or more further batches with respective time steps earlier than the different respective time steps of the one or more other batches before being sent to the application, wherein the ranking constraint is determined based on parity between different groups associated with individual ones of the plurality of data items; and
providing, by the ranking system, the fair ranking of the plurality of data items to the application.
7. The method of claim 6 , wherein the re-ranking policy is trained by a learning to search technique that iteratively selects data items from different queues corresponding to different groups of data items according to a reference policy when re-ranking one or more other batches of data items from a test data set and wherein respective deviations are made from a reference policy when making individual selections from the different queues.
8. The method of claim 6 , wherein the re-ranking policy is trained by a learning to search technique that iteratively selects data items from different queues corresponding to the different groups associated with the individual ones of the plurality of data items according to a reference policy when re-ranking one or more other batches of data items from a test data set and wherein respective deviations are made from a combination of a reference policy and a previously learned policy when making individual selections from the different queues.
9. The method of claim 6 , wherein the re-ranking model applies a deterministic re-ranking policy.
10. The method of claim 9 , wherein the deterministic re-ranking policy finds a most highly ranked data item associated with a first group which is below a data item associated with a second group in the ranking of the batch of items and swaps the data item associated with the second group with the most highly ranked data item in the ranking of the batch of items.
11. The method of claim 6 , wherein the re-ranking policy satisfies a plurality of ranking constraints applicable to the batch of data items and one or more other batches of data items from the stream of data items including the ranking constraint.
12. The method of claim 6 , wherein the ranking constraint is a fairness constraint.
13. The method of claim 12 , wherein the fairness constraint is demographic disparity.
14. One or more non-transitory, computer-readable storage media, storing program instructions that when executed on or across one or more computing devices, cause the one or more computing devices to implement a ranking system providing fairness across a plurality of rankings, comprising:
receiving, by a ranking system, a stream of data items over time;
providing, as part of continuous monitoring of the ranking system, respective fair rankings of respective pluralities of data items for a plurality of time steps including a current time step and one or more time steps earlier than the current time step, wherein providing a fair ranking for the current time step comprises:
performing, by the ranking system, a first ranking of a plurality of data items in a batch of data items obtained from the stream of data items, wherein the batch of data items are associated with the current time step, and wherein the initial ranking is generated using a ranking model that is a machine learning model trained to generate the first ranking of the plurality of data items in the batch in a descending order of relevance for an application;
re-ranking, by the ranking system, the first ranking of the plurality of data items in the batch to generate a second ranking of the plurality of data items in the batch according to a re-ranking policy that satisfies a ranking constraint applicable, in aggregate, to both the batch of data items and one or more other batches of data items from the stream of data items associated with different respective time steps earlier than the current time step, wherein the one or more other batches of data items were previously ranked using the ranking model and re-ranked, using the re-ranking model, with respective rankings of one or more further batches with respective time steps earlier than the different respective time steps of the one or more other batches before being sent to the application, wherein the ranking constraint is determined based on parity between different groups associated with individual ones of the plurality of data items; and
sending, by the ranking system, the re-ranking of the plurality of data items to the application.
15. The one or more non-transitory, computer-readable storage media of claim 14 , wherein the re-ranking policy is trained by a learning to search technique that iteratively selects data items from different queues corresponding to the different groups associated with the individual ones of the plurality of data items according to a reference policy when re-ranking one or more other batches of data items from a test data set and wherein respective deviations are made from a reference policy when making individual selections from the different queues.
16. The one or more non-transitory, computer-readable storage media of claim 14 , wherein the re-ranking policy is trained by a learning to search technique that iteratively selects data items from different queues corresponding to different groups associated with the individual ones of the plurality of data items according to a reference policy when re-ranking one or more other batches of data items from a test data set and wherein respective deviations are made from a combination of a reference policy and a previously learned policy when making individual selections from the different queues.
17. The one or more non-transitory, computer-readable storage media of claim 14 , wherein the re-ranking policy is a deterministic re-ranking policy.
18. The one or more non-transitory, computer-readable storage media of claim 17 , wherein the deterministic re-ranking policy finds a most highly ranked data item associated with a first group which is below a data item associated with a second group in the ranking of the batch of items and swaps the data item associated with the second group with the most highly ranked data item in the ranking of the batch of items.
19. The one or more non-transitory, computer-readable storage media of claim 14 , wherein the re-ranking policy satisfies a plurality of ranking constraints applicable to the batch of data items and one or more other batches of data items from the stream of data items including the ranking constraint.
20. The one or more non-transitory, computer-readable storage media of claim 14 , wherein the ranking constraint is a fairness constraint.