Pupil-Based Prediction of Affect in VR: A Machine Learning Approach

Adapting VR scenarios in real time based on users’ emotional responses is a promising advancement for developers. However, most VR environments are either "blind" to emotions or depend on intrusive, contact-based methods. This study introduces a novel, noninvasive approach for predicting arousal and valence using pupil reactivity and gaze features. Data from 105 participants (ages 18–30) viewing 120 emotional images via a VR headset with eye tracking informed machine learning models. These models, leveraging features like mean and standard deviation of pupil data, achieved F1 scores of 0.64 (valence) and 0.73 (arousal), highlighting potential real-world applications in VR development.