Affective Video Recommender System

Video recommendation is the task of providing users with customized media content conventionally done by considering historical user ratings. We develop classifiers that learn from human faces toward a video recommender system that utilizes displayed emotional reactions to previously seen videos for predicting preferences. We use a dataset collected from subjects who watched videos selected to elicit different emotions, to model two related problems: (1) prediction of user rating and (2) whether a user would recommend a particular video. The classifiers are trained on two forms of face-based features: facial expressions and skin-estimated pulse. In addition, the impact of data augmentation and instance size are studied.