VisArena: A testbed for visualization with synthetic data tables

Siwei Fu, Junjie Jin, Qi Ma, Zhongsu Luo, Kai Xiong, Yingcai Wu

Abstract: Evaluation appears in every corner of the visualization community. Researchers and engineers usually need to create synthetic data tables to validate and evaluate the characteristics of visualization techniques. However, due to the large space and interaction of data characteristics, generating data with various requirements is tedious and challenging. Situated in the domain of multivariate data visualization, we present a design study to develop VisArena, a system for visualization evaluation with synthetic data tables. Based on prompt engineering for generative pre-trained transformers, VisArena features a data generation pipeline that is extensible in the number of constraints. At the same time, to ensure reliability, the system borrows a satisfiability modulo theories (SMT) solver to ensure that all constraints are satisfied in the resulting data tables. For better usability, the interface of VisArena is designed in an iterative manner with three visualization practitioners and offers the ability to specify data attributes, inspect visualization results, and customize evaluation metrics. To demonstrate the effectiveness of VisArena, we report an in-depth case study with expert users. Further, we show how VisArena can serve as a testbed for hierarchical data visualization with an algorithmic study.