Shared per content provider prediction models
An online system, such as a social networking system, generates shared models for one or more clusters of categories. A shared model for a cluster is common to the categories assigned to the cluster. In this manner, the shared models are specific to the group of categories (e.g., selected content providers) in each cluster while requiring a reasonable computational complexity for the online system. The categories are clustered based on the performance of a model specific to a category on data for other categories.
1. A method comprising:
maintaining a plurality of machine-learned models for a plurality of clusters;
generating the plurality of clusters by clustering one or more content providers based on similarity between performance metrics of the one or more content providers, the performance metrics for at least one content provider indicate performance of a machine-learned model, for the at least one content provider, on validation datasets of other content providers;
assigning a content provider, among the one or more content providers, to a corresponding cluster based on performance of a machine-learned model, for the corresponding cluster, on data associated with the content provider;
receiving a request for predicting user responses to a content item associated with the content provider;
selecting a machine-learned model corresponding to a cluster assigned to the content provider; and
applying the selected machine-learned model to the content item to generate predicted user responses to the content item.
2. The method of claim 1 , wherein the machine-learned model for the cluster is trained on datasets of content providers in the one or more content providers assigned to the cluster.
3. The method of claim 1 , further comprising generating the plurality of machine-learned models for the plurality of clusters to reduce a loss function.
4. The method of claim 1 , wherein the content provider is assigned to the corresponding cluster based on characteristics common to the content provider and the corresponding cluster.
5. The method of claim 3 , wherein the loss function indicates predictive error of the plurality of machine-learned models, and indicates computational complexity of the plurality of machine-learned models that increases as a number of the plurality of machine-learned models increases.
6. The method of claim 5 , further comprising re-training the machine-learned model for the corresponding cluster using datasets of the content provider assigned to the cluster.
7. The method of claim 1 , further comprising maintaining a general machine-learned model for the plurality of clusters.
8. The method of claim 7 , further comprising:
receiving a request for predicting user responses to another content item associated with another content provider; and
responsive to determining that the another content provider is not assigned to a cluster, applying the general machine-learned model to the another content item to generate predicted user responses to the another content item.
9. The method of claim 7 , further comprising:
responsive to receiving an indication of failure in the selected machine-learned model, applying the general machine-learned model to the content item to generate predicted user responses to the content item.
10. The method of claim 1 , further comprising:
querying at least one database associated with the plurality of machine-learned models to identify at least one machine-learned model for the content provider.
11. A non-transitory computer-readable medium storing instructions for execution by at least one processor, the instructions when executed by the at least one processor cause:
maintaining a plurality of machine-learned models for a plurality of clusters;
generating the plurality of clusters by clustering one or more content providers based on similarity between performance metrics of the one or more content providers, the performance metrics for at least one content provider indicate performance of a machine-learned model, for the at least one content provider, on validation datasets of other content providers;
assigning a content provider, among the one or more content providers, to a corresponding cluster based on performance of a machine-learned model, for the corresponding cluster, on data associated with the content provider;
receiving a request for predicting user responses to a content item associated with the content provider;
selecting a machine-learned model corresponding to a cluster assigned to the content provider; and
applying the selected machine-learned model to the content item to generate predicted user responses to the content item.
12. The non-transitory computer-readable medium of claim 11 , wherein the machine-learned model for the cluster is trained on datasets of content providers in the one or more content providers assigned to the cluster.
13. The non-transitory computer-readable medium of claim 11 , wherein the instructions, when executed, further cause generating the plurality of machine-learned models for the plurality of clusters to reduce a loss function.
14. The non-transitory computer-readable medium of claim 11 , wherein the content provider is assigned to the corresponding cluster based on characteristics common to the content provider and the corresponding cluster.
15. The non-transitory computer-readable medium of claim 13 , wherein the loss function indicates predictive error of the plurality of machine-learned models, and indicates computational complexity of the plurality of machine-learned models that increases as a number of the plurality of machine-learned models increases.
16. The non-transitory computer-readable medium of claim 15 , wherein the instructions, when executed, further cause re-training the machine-learned model for the corresponding cluster using datasets of the content provider assigned to the cluster.
17. The non-transitory computer-readable medium of claim 11 , wherein the instructions, when executed, further cause maintaining a general machine-learned model for the plurality of clusters.
18. The non-transitory computer-readable medium of claim 17 , wherein the instructions, when executed, further cause:
receiving a request for predicting user responses to another content item associated with another content provider; and
responsive to determining that the another content provider is not assigned to a cluster, applying the general machine-learned model to the another content item to generate predicted user responses to the another content item.
19. The non-transitory computer-readable medium of claim 17 , wherein the instructions, when executed, further cause:
responsive to receiving an indication of failure in the selected machine-learned model, applying the general machine-learned model to the content item to generate predicted user responses to the content item.
20. The non-transitory computer-readable medium of claim 11 , wherein the instructions, when executed, further cause:
querying at least one database associated with the plurality of machine-learned models to identify at least one machine-learned model for the content provider.