Machine learning classification of user attributes via eye movements

Abstract

The advent of modern eye tracking devices has spawned a plethora of new research on eye movements. Applications of these research results include the prediction of diseases, of biometrics, of gender, or of cognitive developments in children. One par- ticularly well studied topic is user identification. Another, less well studied one is gender prediction. In this thesis, a common framework to predict users and gen- der is proposed. Using this framework, we were able to improve the state-of-the-art accuracies for both user identification and gender prediction. Further, unlike previ- ous studies, the proposed approach was tested with different datasets consisting of varying stimuli. We identify several factors that affect the identification accuracy. Our main improvements in identification accuracy are due to three factors, select- ing optimal hyper-parameters of the segmentation algorithm, adding higher-order derivatives, and including blink information. For gender prediction, the thesis es- tablishes several new insights. For instance, that gender prediction is possible for prepubescent children aged 9–10. Previous research had suggested that significant gender differences in eye movements can only be observed in adults. Various factors are identified which affect the accuracy of gender prediction; for example, the length of the gaze trajectory, possible fatigue of the participant (gender prediction works better in the presence of fatigue), and the choice of feature ranking algorithms.