Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for using machine learning models to expand user groups while preserving user privacy and data security are described. In one aspect, a method includes receiving, for a web-based resource, a set of user group identifiers for a set of user interest groups that each include, as members, one or more users that requested content from the web-based resource over a given time period. A seed user list that includes user identifiers for at least a portion of the users in the set of user interest groups is created. A similar audience machine learning model is generated based on a set of one or more feature values corresponding to one or more features of the users corresponding to the user identifiers in the seed user list. A set of similar users is identified using the model.