A digital magazine server identifies content items for recommendation to a user based on content items with which the user previously interacted. Based on key phrases and terms in content items with which the user previously interacted, topics are associated with the content items and used to generate a vector for each content item. The vectors are used to generate clusters including one or more content items. A characteristic vector is generated for each cluster based on the vectors generated for content items within a cluster. Candidate content items are retrieved and topics included in the candidate content items are used along with the characteristic vectors to determine a measure of similarity between candidate content items and various clusters. Candidate content items with at least a threshold measure of similarity to a cluster are selected for presentation to the user.