An attention-aware map interaction framework based on eye tracking and generative models

Traditional map interaction methods (such as mouse clicks or touch operations) rely heavily on explicit input, which tends to interrupt users’ spatial cognitive processes and increase their cognitive load. Although eye-tracking technology can capture users’ attention distributions, existing methods generally struggle to infer deep semantic intentions from gaze behaviors and lack the ability to dynamically generate map content on the basis of real-time cognitive states. Therefore, this study proposes a perception-enhanced map interaction framework driven by user cognition and attention, in which gaze behavior is not treated as a direct control mechanism but is instead regarded as a cognitive cue. By quantifying interest values and performing cognitive reasoning with a large language model (LLM), and by dynamically generating map symbols using a LoRA-fine-tuned Stable Diffusion model, a closed-loop workflow from gaze analysis to map symbol rendering is established. A prototype system was developed and validated across three typical scenarios. The results show that the prediction accuracy of areas of interest achieved an average IoU of 0.7 and an F1 score of 0.8, while map interaction efficiency improved by approximately 26.34–36.38%. Compared with traditional mouse interaction methods, the task completion rate increased by approximately 19.7%, the cognitive load was significantly reduced, and the system usability score (SUS) was significantly greater. By understanding users’ attention distribution, this study enhances map visual representations and provides a feasible technical pathway for the transition of map interaction paradigms from passive response to active understanding.