Techniques are provided for selecting which videos to recommend to users by predicting the degree to which recommending each video will satisfy certain goals. To make the predictions, a trained machine learning engine is fed both collaborative filtering parameter values and content-based filtering parameter values. In the case of video-to-video recommendations, the collaborative filtering parameter values may be based on a video pair that includes a video in which a user has already demonstrated an interest. The machine learning engine generates a machine-learning score for each video. The machine learning scores are used as the basis for selecting which videos to recommend to a particular user.