Evaluating the potential of profiling users from visual behavior data in pedestrian navigation

Recognizing user attributes during wayfinding is essential for developing adaptive and personalized navigation aids. Machine learning models trained on gaze data offer a promising approach, yet most prior studies have been conducted in controlled laboratory settings, leaving their applicability and interpretability in dynamic real-world contexts unclear. This study investigates the potential of eye movement data for user profiling in both immersive virtual reality and real-world environments. Using machine learning techniques, we analyzed gaze behavior to recognize five user attributes: gender, expertise, spatial ability, environmental familiarity, and navigation task. A comprehensive leave-one-X-out validation (X = task, person, environment) was adopted to evaluate model generalizability across tasks, users, and environments. The Light Gradient-Boosting Machine (LightGBM) achieved the best overall performance, with task recognition consistently outperforming other attributes. However, generalizability declined for unseen users and environments. Statistical analyses further showed that basic statistical features, pupil dilation measures, and fixation semantic features were particularly informative for user profiling. These findings highlight both the potential and current limitations of gaze-based user modeling for personalized navigation in real-world applications.