An online system selects a number of content items and presents the selected content items through a feed to a target user, where each selected candidate content item is likely to cause the target user to post his/her new content in response to the selected candidate content item within a short period of time. The online system selects the candidate content items for presentation through the feed using a trained post session prediction model. A ranking score for a candidate content item is determined based on a probability value indicating likelihood that the candidate content item causes the target user to post new content. The probability value is determined by applying a trained model to user features of the target user and content features of the candidate content item. The online system ranks the candidate content items based on their ranking scores and present the feed to the target user.