Personal rating prediction for on-line video lectures using gaze information

Automatic prediction of personal preference for lecture video is becoming exceedingly important as the available volume of such content is expanding rapidly. Gaze information is believed to be an convenient and useful indicator of cognitive process due to its non-invasive characteristic. We build up a small-scale dataset of viewer´s gaze during watching TED Talk videos, and propose a set of gaze features to predict personal preference. Besides the conventional gaze statistics features, we also propose new features that analyze the relationship between visual saliency and gaze points. The proposed features are proved to be effective in the evaluation of personal rating prediction and could achieve over 85% accuracy in the leave-one-out test and over 75% in the half-half train-test-split setting. To our best knowledge, this is the first work on rating prediction for video lectures using gaze.