“Contextual Similarity Abstraction Techniques for Spatial Data Analysis”

July 13th, 2017

Categories: Applications, MS / PhD Thesis, Software, Visualization, Visual Analytics, Visual Informatics


EVL PhD Candidate, Timothy Luciani presents his research “Contextual Similarity Abstraction Techniques for Spatial Data Analysis” during his preliminary exam.

Date: Thursday July 13, 2017
Time: 10:00am
Room: ERF 2068 (Cybercommons)

Committee: Dr. G. Elisabeta Marai (Chair), Dr. Robert Kenyon, Dr. Andrew Johnson, Dr. Angus Forbes, Dr. Klaus Mueller (Stony Brook University, NY)

Comparative analysis of modern large-scale datasets is a core task in scientific visualization. This task entails identifying both global similarities among entire data collections, and local similarities between smaller data subsets to help provide critical insights into data trends and potentially lead to classification, validation and discovery. To this day, Shneiderman’s visual information-seeking mantra remains the most common approach to designing comparative visualization applications: Overview first, zoom and filter, then details on demand. However, for large-scale datasets, this mantra becomes infeasible due to the high volume (number of entries and dimensionality) of the data. Van Ham and Perer explore an alternative, context-based model that emphasizes the use of contextual information as a means of data reduction: Search, show context, expand on demand. While powerful, this paradigm has not been well studied outside the domain of large-scale, dense graphs. Specifically, little is known on how well this approach transfers to domain-specific, datasets that are heterogeneous, spatially and statistically dense yet locally sparse, and possess spatial/non-spatial properties.

To help explore this problem space, I propose a set of comparative analysis techniques which extend the Van Ham and Perer paradigm to spatial, scientific data. These techniques employ a blending of derived, quantitative similarity metrics and high-density visual abstractions - implemented through various integrated systems - to handle the inherent global data density and local attribute sparsity found in these datasets. Each of these approaches improves the likelihood of discovering patterns and similarities by first limiting the data to a user-defined subset before providing abstracted, contextual overviews.