Researchers unveil systematic design method for visual multiplexing in ML

A new paper from Yiwen Xing, Afrah Farea, Saiful Khan, and Min Chen tackles the challenge of "visual multiplexing"—designing visualizations that convey multiple pieces of information for the same data point simultaneously. While many existing designs use this approach unconsciously, the authors argue that a systematic exploration of the design space is rare. They introduce a structured method that constrains the vast design space to an application-specific subset, using ML workflows for developing models that approximate partial differential equations (PDEs) as a case study.

In these workflows, researchers frequently need to analyze relationships among multiple 2D scalar fields, but simply superimposing heatmaps is ineffective. The proposed method consists of three main design steps plus a pre-design phase for domain grounding and theoretical analysis. It emphasizes co-design and evaluation with domain experts, leading to identifications of optimal default multiplexing designs and controllable variations via a user interface. This approach promises to make complex multi-field analysis more intuitive and efficient for ML scientists.