“However, our predicted model output will now be in the downsampling space. For instance, consider that if our engagement rate is 5% and we select only 10% negative samples, our average predicted engagement rate will be near 50%.” This statement is not clear. Aren’t the predictions of the model the probability of an ad being viewed by a user? If yes, then why would engagement rate matter and why the calibration? The auction should have access to real engagement rates anyway. Please clarify.