Data Type Agnostic Visual Sensitivity Analysis

Area of Science:

• Computational science and data analysis
• Spatial data analysis
• Scientific visualization

Background:

• Modern science and industry heavily depend on computational models for simulation, prediction, and analysis.
• Spatial blind source separation (SBSS) is a specialized model for spatial data analysis, outperforming non-spatial methods like PCA.
• A significant challenge in SBSS practical application is the complex tuning of two parameters, necessitating parameter space analysis.

Purpose of the Study:

• To address the difficulty of sensitivity analysis (SA) for SBSS due to spatial data inputs and outputs.
• To develop a data-type-agnostic visual analytics prototype for sensitivity analysis tailored to SBSS and similar contexts.
• To create a tool that simplifies parameter tuning by requiring only dissimilarity measures.

Main Methods:

• Developed a visual analytics prototype based on expert consultations for data type agnostic visual sensitivity analysis.
• The prototype requires only dissimilarity measures for parameter settings and outputs, accommodating complex spatial data.
• Evaluated the prototype through heuristic methods with visualization experts and interviews with SBSS experts.

Main Results:

• The prototype successfully facilitated the confirmation of known and suspected parameter-output relationships.
• Study participants identified surprising associations and pinpointed parameter subspaces for future investigation.
• Demonstrated the approach's transferability by applying it to microclimate simulations.

Conclusions:

• The developed visual analytics prototype offers a practical solution for sensitivity analysis in spatial blind source separation.
• The approach's reliance on dissimilarity measures makes it adaptable to various contexts with complex spatial data.
• The study highlights future research opportunities in advanced sensitivity analysis for computational models.