Recommendations to promote content discovery

1 . A computer-implemented method, comprising:

identifying candidate content items from a set of eligible content items;

assigning a corresponding rank to each of the candidate content items using an objective function that mitigates a popularity bias among the candidate content items, wherein the objective function comprises an inverse dominance term to mitigate the popularity bias among the candidate content items, and wherein the inverse dominance term is 1 divided by a logarithm of a number of impressions corresponding to a subset of the candidate content items recorded during a time period;

inputting first noise into the objective function which is included in a recommender system of a computing device, wherein the first noise is inputted into the objective function using dithering, and wherein inputting the first noise promotes less bias in the ranked candidate content items by increasing an ability of the computing device to learn content and impression incrementally:

determining an impression-distribution mix of the ranked candidate content items; and

causing one or more of the ranked candidate content items to be displayed based on the impression-distribution mix.

2 . The computer-implemented method of claim 1 , wherein the objective function further comprises a parameter value that controls an amount of mitigation for the popularity bias.

3 . The computer-implemented method of claim 1 , wherein determining the impression-distribution mix comprises determining the impression-distribution mix using a proportional-integral-derivative (PID) controller.

4 . The computer-implemented method of claim 1 , wherein the candidate content items are identified using a plurality of candidate generators.

5 . The computer-implemented method of claim 4 , wherein the plurality of candidate generators includes two or more of a retention-based candidate generator, a semantic-similarity candidate generator, an engagement-similarity candidate generator, a behavior-similarity candidate generator, or user-data candidate generator.

6 . The computer-implemented method of claim 4 , further comprising:

inputting second noise into the plurality of candidate generators which are included in the recommender system of the computing device, wherein the second noise is inputted into the plurality of candidate generators using dithering, wherein inputting the second noise promotes a discovery capability of the computing device by increasing a likelihood of the computing device discovering content items that otherwise would not be included in the set of eligible content items, and wherein a number of content items in the set of eligible content items is higher with the second noise input than without the second noise input.

7 . A non-transitory computer-readable medium with instructions stored thereon that, when executed by one or more hardware processors, cause the one or more hardware processors to perform operations comprising:

identifying candidate content items from a set of eligible content items;

assigning a corresponding rank to each of the candidate content items using an objective function that mitigates a popularity bias among the candidate content items, wherein the objective function comprises an inverse dominance term to mitigate the popularity bias among the candidate content items, and wherein the inverse dominance term is 1 divided by a logarithm of a number of impressions corresponding to a subset of the candidate content items recorded during a time period;

inputting first noise into the objective function which is included in a recommender system of a computing device having the one or more hardware processors, wherein the first noise is inputted into the objective function using dithering, and wherein inputting the first noise promotes less bias in the ranked candidate content items by increasing an ability of the computing device to learn content and impression incrementally;

determining an impression-distribution mix of the ranked candidate content items; and

causing one or more of the ranked candidate content items to be displayed based on the impression-distribution mix.

8 . The non-transitory computer-readable medium of claim 7 , wherein the objective function further comprises a parameter value that controls an amount of mitigation for the popularity bias.

9 . The non-transitory computer-readable medium of claim 7 , wherein determining the impression-distribution mix comprises determining the impression-distribution mix using a proportional-integral-derivative (PID) controller.

10 . The non-transitory computer-readable medium of claim 7 , wherein the candidate content items are identified using a plurality of candidate generators.

11 . The non-transitory computer-readable medium of claim 10 , wherein the plurality of candidate generators includes two or more of a retention-based candidate generator, a semantic-similarity candidate generator, an engagement-similarity candidate generator, a behavior-similarity candidate generator, or user-data candidate generator.

12 . The non-transitory computer-readable medium of claim 10 , wherein the operations further comprise:

inputting second noise into the plurality of candidate generators which are included in the recommender system of the computing device, wherein the second noise is inputted into the plurality of candidate generators using dithering, wherein inputting the second noise promotes a discovery capability of the computing device by increasing a likelihood of the computing device discovering content items that otherwise would not be included in the set of eligible content items, and wherein a number of content items in the set of eligible content items is higher with the second noise input than without the second noise input.

13 . A computing device, comprising:

one or more hardware processors; and

a non-transitory computer readable medium coupled to the one or more hardware processors, with instructions stored thereon, that when executed by the one or more hardware processors cause the one or more hardware processors to perform operations comprising:

identifying candidate content items from a set of eligible content items;

assigning a corresponding rank to each of the candidate content items using an objective function that mitigates a popularity bias among the candidate content items, wherein the objective function comprises an inverse dominance term to mitigate the popularity bias among the candidate content items, and wherein the inverse dominance term is 1 divided by a logarithm of a number of impressions corresponding to a subset of the candidate content items recorded during a time period;

inputting noise into the objective function which is included in a recommender system of the computing device, wherein the noise is inputted into the objective function using dithering, and wherein inputting the noise promotes less bias in the ranked candidate content items by increasing an ability of the computing device to learn content and impression incrementally;

determining an impression-distribution mix of the ranked candidate content items; and

causing one or more of the ranked candidate content items to be displayed based on the impression-distribution mix.

14 . The computing device of claim 13 , wherein the objective function further comprises a parameter value that controls an amount of mitigation for the popularity bias.

15 . The computing device of claim 13 , wherein determining the impression-distribution mix comprises determining the impression-distribution mix using a proportional-integral-derivative (PID) controller.

16 . The computing device of claim 13 , wherein the candidate content items are identified using a plurality of candidate generators.

17 . The computing device of claim 16 , wherein the plurality of candidate generators includes two or more of a retention-based candidate generator, a semantic-similarity candidate generator, an engagement-similarity candidate generator, a behavior-similarity candidate generator, or user-data candidate generator.