Revealing Data Structures by Creating Networks Through Distances. Visual Analytics Approaches for Exploring Multidimensional Datasets

In Visual Analytics (VA), we combine data mining and other analytical techniques with data visualization in order to gain insights from data. As the size and complexity of data increase, the reasoning process becomes more arduous and challenging for analysts. This thesis aims to enhance the identification of patterns using flexible methods in VA. It centers around the research question: 'How to increase the analyst's ability in VA systems to identify patterns while data context is preserved?'. The main contribution of this thesis is the employment of network visualization to depict relationships between multidimensional elements based on the similarity between datapoints. This abstract representation can make complex relationships visually accessible to the user. To illustrate the effectiveness of network visualization for data analysis, we present methods for the two principal data summarization tasks, clustering and dimension reduction.Data clustering often requires looking at different resolutions in data, especially when data is heterogeneous. However, by inspecting them as similarity networks, we can find a scalable environment to navigate through different granularities of data. The classical visualization of dimensionality reduction results relies on scatterplots, reducing a high number of dimensions to two or three. Using networks such as topological data analysis, we can reduce dimensionality to a space in which the distance between points is defined by the path length in the network. The layout or visualization of the network can then be more dynamic because the connectivity of the nodes is preserved. With this type of representations, analysts can explore and recognize global and local signals in data, depending on the task at hand.In addition to these general techniques, we describe applications in which topological data analysis contributes to exploring and evaluating complex data. The present thesis systematically discusses the benefits and challenges of the proposed solutions and outlines future research directions.

Publication year:2021

Accessibility:Closed