Abstract: A system may receive user information associated with a user, third-party information associated with a third party, and a relevance-ordered list of media carousels comprising a plurality of media carousels. A system may collect historical log information comprising a plurality of displayed pages comprising media carousels and assign a reward value to each of the displayed pages to generate a reward vector, the reward value based on a user interaction associated with each displayed page. A system may estimate a logging policy based in part on the historical log information, and a target policy based in part on the reward vector. A system may train a carousel selection model by the target policy and the logging policy, then use the carousel selection model to generate a result. A system may select a plurality of media carousels from the relevance-ordered list.

Abstract: Embodiments of a content recommendation or selection system are described. The system uses a pipeline of machine learning models to select content for a user. In embodiments, a first user model generates a first score of content categories for the user based on short-term user data. A second user model generates a second score of the categories for the user based on long-term user data. The two scores are combined to select the categories to include on the content user interface. In embodiments, new categories are added to the recommendations based on an exploration-exploitation algorithm. In embodiments, content categories are organized on the user interface in a manner to promote neighborhood diversity. Advantageously, the machine learning pipeline enables independent configurability of various objectives of the content recommendation or selection system and reduces the amount of machine learning resources needed to implement the system.