A few of the prior works have leveraged provenance data to understand the visualization system itself and to evaluate its usefulness. Here, it is important to distinguish between conducting statistical analysis on coarse user study metrics (e.g., speed, accuracy, and preference) and the non-trivial analysis of provenance data for the primary purpose of evaluating a visualization design or system. For instance, Bylinskii et al. trained a neural network on mouse click data to create an automated model that learns the relative importance of visual elements for a given design. Smuc et al. captured the provenance to identify when users have insights. Gomez and Laidlaw modeled task performance on crowd workers to evaluate system design and help guide encoding choices. Blascheck et al. created a visual analytics system for evaluating an interactive visualization system. Among other techniques, they used pattern matching methods to uncover similarities within the provenance data of multiple users.