Shape-based visual analysis of multidimensional projections

• Grant number: 14/01692-6
• Support Opportunities: Scholarships abroad - Research Internship - Doctorate
• Effective date (Start): June 01, 2014
• Effective date (End): May 31, 2015
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Rosane Minghim
• Grantee: Renato Rodrigues Oliveira da Silva
• Supervisor: Alexandru C. Telea
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• Research place: University of Groningen, Netherlands
• Associated to the scholarship: 11/18838-5 - Shape Based Information Visualization, BP.DR

Abstract

Projection techniques provide a way to quickly find and analyze useful patterns in multidimensional data. Their goal is to reduce the dimensionality of the original data, usually projecting them to 2 dimensions and keeping the original distance structure as much as possible. A projection is usually visualized on the screen space as a 2D point-based layout (scatterplot) that encodes similarity relationship in the original multidimensional data. This visual encoding, albeit informative, is intended to describe properly only the relationship of pairs of elements, not the relationship for the whole data set. As the size of the analyzed data grows, it is important to provide a way to visually summarize and explore it as groups rather than individual elements. Usually the scatterplot approach is not easy to analyze, and it may be a challenging task to identify groups of related elements.This project aims at providing visual techniques to allow an efficient analysis of multidimensional embedding laid out as scatterplots. Based on the points layout, we will create continuous and smooth shapes that describe the locations of instance groups and explaining their proximity in the layout. Then additional visual features can be rendered on each shape, to describe additional information and highlight the most important groups that met the users exploration expectations. This approach is intended to be multiscale, letting the user control the process of shapes creation by a continuous parameter, and visualize several levels of detail of the same data set. This flexibility will allow the creation of multiple shape visualization configurations, which can be combined and grant a richer visual exploration to the user. (AU)