Toward implicit gaze-based map interaction: predicting selection intentions from eye movements

Gaze-based interaction provides a natural and hands-free alternative for map use but often depends on explicit methods such as dwell time, which may disrupt typical eye movement behavior. To enable more seamless interaction, this study explores implicit prediction of user intention from eye movement data using machine learning. An experiment was conducted under both gaze-based and mouse-based modes. Selection intentions were predicted using a light gradient boosting machine (LightGBM) classifier applied to backward sliding time windows. Results show that intentions can be reliably identified from eye movement patterns 500–2000 ms before interaction, achieving accuracies of 86.4% in gaze mode and 85.8% in mouse mode with a 1000-ms window and 1000-ms step. Cross-dataset and leave-one-group-out validations further confirmed generalizability, with accuracies exceeding 80%. Feature importance analysis revealed consistent patterns of longer fixation durations, longer gaze-target distance, and reduced saccade rate preceding interactions. These modality- and participant-independent features suggest that gaze behavior encodes a robust cognitive process underlying map selection. These findings demonstrate the feasibility of intention prediction in map-reading tasks and highlight the potential of integrating real-time gaze analysis into adaptive gaze-based map systems.