Described herein are methods for training a machine learning-based predictor of affective response to stimuli. The methods involve receiving samples comprising temporal windows of token instances to which a user was exposed, and target values representing affective response annotations of the user in response to the temporal windows of token instances. This data is used for the training of the predictor along with values indicative of the number of the token instances in the temporal windows of token instances, which are used to compensate for non-linear effects resulting from saturation of the user.