Visual Analytics for Blind Source Separation in Time and Space

Abstract

Multivariate data are often collected over time or space, e.g., in the environmental sciences, healthcare, or commerce. Analysts are interested in underlying processes and signals in the datasets or producing models for downstream tasks, such as interpolation, forecasting, or prediction. These goals can be difficult to achieve: When signals are weak and masked by others, plotting the data does not show them. Multivariate spatial modeling requires considering interactions between variables, which grow with the square of the dataset’s dimensionality. Blind Source Separation (BSS) is a latent variable model for multivariate temporal or spatial data and can aid those goals. However, applying it to a dataset can be challenging. Among others, the tuning parameter space is large and intricate, and exploring it leads to a wealth of latent variables to consider. In this thesis, we propose visual analytics approaches, i.e., automated data mining techniques combined with interactive visualizations, to tackle BSS’s challenges. We follow established visualization research methodology, such as the Design Study Methodology or the Nested Model. As a result, we suggest visual parameter optimization for Temporal and Spatial BSS, as well as visual sensitivity analysis for Spatial BSS. We embed our contributions in the visualization literature, discuss their relevance to and advancement of our research questions, and outline possible future research directions.