Chartist: Task-driven Eye Movement Control for Chart Reading

• DOI: 10.48550/arxiv.2502.03575
• Publication Date: 2025-02-05
• Authors: Danqing Shi; Yao Wang; Yunpeng Bai; Andreas Bulling; Antti Oulasvirta
• Keywords: Movement control

To design data visualizations that are easy to comprehend, we need understand how people with different interests read them. Computational models of predicting scanpaths on charts could complement empirical studies by offering estimates user performance inexpensively; however, previous have been limited gaze patterns and overlooked the effects tasks. Here, contribute Chartist, a computational model simulates users move their eyes extract information from chart in order perform analysis tasks, including value retrieval, filtering, finding extremes. The novel contribution lies two-level hierarchical control architecture. At high level, uses LLMs comprehend gained so far applies this representation select goal for lower-level controllers, which, turn, accordance sampling policy learned via reinforcement learning. is capable human-like task-driven across various It can be applied fields such as explainable AI, visualization evaluation, optimization. While it displays limitations terms generalizability accuracy, takes modeling promising direction, toward understanding human behaviors interacting charts.