Computer Science and Engineering, Department of


First Advisor

Jieting Wu

Date of this Version

Fall 11-30-2018


A DISSERTATION Presented to the Faculty of The Graduate College at the University of Nebraska In Partial Fulfilment of Requirements For the Degree of Doctor of Philosophy, Major: Computer Science, Under the Supervision of Professor Hongfeng Yu. Lincoln, Nebraska: November, 2018

Copyright (c) 2018 Jieting Wu


A tremendous increase in the scale of graphs has been witnessed in a wide range of fields, which demands efficient and effective visualization techniques to assist users in better understandings of large graphs. Conventional node-link diagrams are often used to visualize graphs, whereas excessive edge crossings can easily incur severe visual clutter in the node-link diagram of a large graph. Edge bundling can effectively remedy visual clutter and reveal high-level graph structures. Although significant efforts have been devoted to developing edge bundling, three challenging problems remain. First, edge bundling techniques are often computationally expensive and are not easy to deploy for web-based applications. The state-of-the-art edge bundling methods often require special system supports and techniques such as high-end GPU acceleration for large graphs, which makes these methods less portable, especially for ubiquitous mobile devices. Second, the quantitative quality of edge bundling results is barely assessed in the literature. Currently, the comparison of edge bundling mainly focuses on computational performance and perceptual results. Third, although the family of edge bundling techniques has a rich set of bundling layout, there is a lack of a generic method to generate different styles of edge bundling.

In this research, I aim to address these problems and have made the following contributions. First, I provide an efficient framework to deploy edge bundling for web-based platforms by exploiting standard graphics hardware functions and libraries. My framework can generate high-quality edge bundling results on web-based platforms, and achieve a speedup of 50X compared to the previous state-of-the-art edge bundling method on a graph with half of a million edges. Second, I propose a new moving least squares based approach to lower the algorithm complexity of edge bundling. In addition, my approach can generate better bundling results compared to other methods based on a quality metric. Third, I provide an information-theoretic metric to evaluate the edge bundling methods. I leverage information theory in this metric. With my information-theoretic metric, domain users can choose appropriate edge bundling methods with proper parameters for their applications. Last but not least, I present a deep learning framework for edge bundling visualizations. Through a training process that learns the results of a specific edge bundling method, my deep learning framework can infer the final layout of the edge bundling method. My deep learning framework is a generic framework that can generate the corresponding results of different edge bundling methods.

Adviser: Hongfeng Yu