Towards the Usability of Dimensionality Reduction for Visual Analytics

Abstract

PhD Defense:

The generation of complex data is increasing rapidly. Data visualization helps users make sense of this data and extract insight. Many methods exist for visualizing low-dimensional data, but high-dimensional data often requires dimensionality reduction. Dimensionality reduction methods allow users to visualize complex data effectively, frequently utilizing as few as two dimensions. However, these techniques are complicated and require careful tuning and understanding. In this talk, I introduce novel approaches to constructing arbitrary projections and tools for better understanding and building trust in said projections. I begin with a discussion of an interview study showcasing the struggles facing practitioners who need to understand and communicate complex data. I then explain two novel algorithms for augmented projection, HyperNP and Unprojection. These methods leverage recent advancements in machine learning to help users tune and correctly understand their projections. I conclude with a discussion of future directions.