In some implementations, a computing device can improve diversity in media item recommendations. For example, the computing device can generate a satisfaction score representing an estimate of the amount satisfaction a user will experience when a candidate media item selected by the user is presented (e.g., played) by the system. The system can determine the difference in satisfaction scores for similar media items over time to determine the amount of satisfaction gained with a subsequent selection and presentation of similar candidate media items. The system can rank candidate media items based the satisfaction gain calculated for each candidate media item and select suggested media items to present to the user based on the candidate media items having the highest satisfaction gain. In some implementations, the computing device can generate satisfaction scores based on an adventurousness score calculated for the user.