Visualizing multidimensional data similarities: improvements and applications

Author(s)

Renato Rodrigues Oliveira da Silva

• Document type: Doctoral Thesis
• Press: São Carlos.
• Institution: Universidade de São Paulo (USP). Instituto de Ciências Matemáticas e de Computação (ICMC/SB)
• Defense date: 2016-12-05

Examining board members

Rosane Minghim; Michael Biehl; Carla Maria Dal Sasso Freitas; Maria Cristina Ferreira de Oliveira; Alexandru Cristian Telea; Marcelo Knorich Zuffo

• Advisor: Rosane Minghim

Abstract

Multidimensional datasetsare increasingly more prominent and important in data science and many application domains. Such datasets typically consist of a large set of observations, or data points, each which is described by several measurements, or dimensions. During the design of techniques and tools to process such datasets, a key component is to gather insights into their structure and patterns, a goal which is targeted by multidimensional visualization methods. Structures and patterns of high-dimensional data can be described, at a core level, by the notion of similarity of observations. Hence, to visualize such patterns, we need effective and efficient ways to depict similarity relations between a large number of observations, each having a potentially large number of dimensions. Within the realm of multidimensional visualization methods, two classes of techniques exist projections and similarity trees which effectively capture similarity patterns and also scale well to the number of observations and dimensions of the data. However, while such techniques show similarity patterns, understanding and interpreting these patterns in terms of the original data dimensions is still hard. This thesis addresses the development of visual explanatory techniques for the easy interpretation of similarity patterns present in multidimensional projections and similarity trees, by several contributions. First, we proposemethodsthat make the computation of similarity treesefficient for large datasets, and also allow their visual explanation on a multiscale, or several levels of detail. We also propose ways to construct simplified representations of similarity trees, thereby extending their visual scalability even further. Secondly, we propose methods for the visual explanation of multidimensional projections in terms of automatically detected groups of related observations which are also automatically annotated in terms of their similarity in the high-dimensional data space. We show next how these explanatory mechanismscan be adapted to handle both static and time-dependent multidimensional datasets. Our proposed techniques are designed to be easy to use, work nearly automatically, handle any typesof quantitativemultidimensional datasets and multidimensional projection techniques, and are demonstrated on a variety of real-world large datasets obtained from image collections, text archives, scientific measurements, and software engineeering. (AU)

• FAPESP's process: 11/18838-5 - Shape Based Information Visualization
• Grantee: Renato Rodrigues Oliveira da Silva
• Support Opportunities: Scholarships in Brazil - Doctorate