Abstract: An online system receives multiple candidate components for including in content items to be presented to online system users. Upon identifying an opportunity to present content to a subject user of the online system, the online system dynamically generates an optimal content item for presentation to the subject user that includes one or more candidate components. Candidate components included in the optimal content item are selected by predicting an affinity of the subject user for each candidate component. The affinity of the subject user for a candidate component may be predicted using a machine-learned model that is trained using historical performance information about content items including the candidate component that were presented to viewing users of the online system having at least a threshold measure of similarity to the subject user. Components of content items used to train the model may be selected using a heuristic (e.g., Thompson sampling).

Abstract: An online system receives multiple candidate components for including in content items to be presented to online system users. Upon identifying an opportunity to present content to a subject user of the online system, the online system dynamically generates an optimal content item for presentation to the subject user that includes one or more candidate components. Candidate components included in the optimal content item are selected by predicting an affinity of the subject user for each candidate component. The affinity of the subject user for a candidate component may be predicted using a machine-learned model that is trained using historical performance information about content items including the candidate component that were presented to viewing users of the online system having at least a threshold measure of similarity to the subject user. Components of content items used to train the model may be selected using a heuristic (e.g., Thompson sampling).