Data visualizations play an essential role in allowing analysts to quickly understand the trends, outliers, and patterns of their data. However, the process of designing the "best" visualizations for a given dataset becomes more complicated. Many factors need to be considered, such as the size of data, type of data, the target analysis task being supported, and even how the visualization needs to be personalized to the audience. Visualization recommendation systems are being proposed in response to reduce user effort in generating visualizations. However, existing visualization recommendation algorithms are rarely evaluated and compared for their performance. Lack of formal evaluation makes it difficult to ascertain whether the newly developed algorithms provide more measurable benefits than the old ones and to identify ways to improve human task performance while designing new visualization recommendation systems.

In this dissertation, we aim to develop a performance-driven visualization recommendation system that can gauge user intentions to recommend effective visualizations--enabling users to derive exciting insights. Towards this goal, we first develop an evaluation-focused framework to enable more effective theoretical and empirical comparisons of visualization recommendation systems. We find that while the recommendation process relies heavily on the knowledge of visualization comparison, current algorithms behave essentially the same in terms of user performance with different analysis tasks. This finding highlights the need to investigate the potential causes for the limited performance benefits of new algorithms. We then assess whether the existing perceptual theory and experiment literature provide enough knowledge in visualization comparison to inform improvements in recommendation systems. We present a literature review comparing and ranking the quality of visualization designs in visual perception and task performance. Through the literature review, we contribute a comprehensive schema to record the specification of visualization designs and the theoretical and experimental ranking of the designs based on various quality metrics and analysis tasks. Although many visualization design pairings have yet to be assessed in the current literature, we can still derive helpful knowledge to inform visualization recommendation systems. We plan to extend our previous work as the next steps to (1) integrate the visualization comparison knowledge from different research works into a single ranking strategy and (2) design an objective description language to describe user's analysis objectives, then finally (3) apply the new ranking strategy in designing a new recommendation system tailored to specific analysis objectives.

Examining Committee: 

 

                          Chair:               Dr. Leilani Battle        
                          Dept rep:         Dr. Huaishu Peng
                          Members:          Dr. Niklas Elmqvist