Ranking user comments on media using reinforcement learning optimizing for session dwell time

1. A method, comprising:

storing comments generated in response to a content item served over a network;

analyzing the comments to determine features associated with each of the comments;

using a scoring model to score each comment based on corresponding features of the comment, wherein an objective associated with the scoring model is changed without re-training the scoring model;

receiving a request to serve a subset of the comments;

determining a plurality of possible rankings of the comments associated with a plurality of possible permutations;

responsive to the request, selecting a ranking of the comments that is one permutation from the plurality of possible rankings of the comments, wherein selecting the ranking is in accordance with a probability distribution of the plurality of possible rankings that is based on scores of the comments;

serving one or more comments identified by the selected ranking over the network to a client device;

determining a dwell time on the one or more comments; and

applying the dwell time to update the scoring model.

2. The method of claim 1 , wherein each permutation from the plurality of possible rankings defines a unique order for at least some of the comments.

3. The method of claim 1 , wherein the features include statistics based on at least one of replies, upvotes, downvotes, or age.

4. The method of claim 1 , wherein the scoring model includes a neural network.

5. The method of claim 4 , wherein updating the scoring model includes adjusting one or more weights of the neural network.

6. The method of claim 1 , wherein updating the scoring model is configured to adjust the scoring model to maximize one or more dwell times of comments ordered according to one or more scores as determined by the scoring model.

7. The method of claim 1 , wherein the scoring model is configured to provide for a given comment a score indicating a relative contribution of the given comment to an optimized ranking.

8. The method of claim 1 , wherein the probability distribution is defined from probabilities of permutations of the plurality of possible rankings, such that a given ranking prioritizing comments predicted to have greater dwell time has a higher probability than a given ranking prioritizing comments predicted to have a lower dwell time, as determined from the scoring of each comment.

9. The method of claim 1 , wherein the dwell time defines a reward for reinforcement learning of the scoring model.

10. The method of claim 1 , wherein the request is generated from an access to the content item.

11. The method of claim 1 , wherein selecting the ranking includes sequentially defining a multinomial distribution and sampling from the multinomial distribution.

12. The method of claim 1 , wherein determining the dwell time includes tracking a quantity of time that the one or more comments are presented at the client device.

13. A non-transitory computer readable medium having program instructions embodied thereon, the program instructions being configured, when executed by a computing device, to cause the computing device to perform operations comprising:

storing comments generated in response to a content item served over a network;

analyzing the comments to determine features associated with each of the comments;

using a scoring model to score each comment based on corresponding features of the comment, wherein an objective associated with the scoring model is changed without re-training the scoring model;

receiving a request to serve a subset of the comments;

determining a plurality of possible rankings of the comments associated with a plurality of possible permutations;

responsive to the request, selecting a ranking of the comments that is one permutation from the plurality of possible rankings of the comments, wherein selecting the ranking is in accordance with a probability distribution of the plurality of possible rankings that is based on scores of the comments;

serving one or more comments identified by the selected ranking over the network to a client device;

determining a dwell time on the one or more comments; and

applying the dwell time to update the scoring model.

14. The non-transitory computer readable medium of claim 13 , wherein each permutation from the plurality of possible rankings defines a unique order for at least some of the comments.

15. The non-transitory computer readable medium of claim 13 , wherein the features include statistics based on at least one of replies, upvotes, downvotes, or age.

16. The non-transitory computer readable medium of claim 13 , wherein the scoring model includes a neural network.

17. The non-transitory computer readable medium of claim 16 , wherein updating the scoring model includes adjusting one or more weights of the neural network.

18. The non-transitory computer readable medium of claim 13 , wherein updating the scoring model is configured to adjust the scoring model to maximize one or more dwell times of comments ordered according to one or more scores as determined by the scoring model.

19. The non-transitory computer readable medium of claim 13 , wherein the scoring model is configured to provide for a given comment a score indicating a relative contribution of the given comment to an optimized ranking.

20. A system comprising at least one server computer, the at least one server computer having:

logic for storing comments generated in response to a content item served over a network;

logic for analyzing the comments to determine features associated with each of the comments;

logic for using a scoring model to score each comment based on corresponding features of the comment, wherein an objective associated with the scoring model is changed without re-training the scoring model;

logic for receiving a request to serve a subset of the comments;

logic for determining a plurality of possible rankings of the comments associated with a plurality of possible permutations;

logic for, responsive to the request, selecting a ranking of the comments that is one permutation from the plurality of possible rankings of the comments, wherein selecting the ranking is in accordance with a probability distribution of the plurality of possible rankings that is based on scores of the comments;

logic for serving one or more comments identified by the selected ranking over the network to a client device;

logic for determining a dwell time on the one or more comments; and

logic for applying the dwell time to update the scoring model.