Understanding the bi-directional relationship between analytical processes and interactive visualization systems.

摘要

Interactive visualizations leverage the human visual and reasoning systems to increase the scale of information with which we can effectively work, therefore improving our ability to explore and analyze large amounts of data. Interactive visualizations are often designed with target domains in mind, such as analyzing unstructured textual information, which is a main thrust in this dissertation.;Since each domain has its own existing procedures of analyzing data, a good start to a well-designed interactive visualization system is to understand the domain experts' workflow and analysis processes. This dissertation recasts the importance of understanding domain users' analysis processes and incorporating such understanding into the design of interactive visualization systems.;To meet this aim, I first introduce considerations guiding the gathering of general and domain-specific analysis processes in text analytics. Two interactive visualization systems are designed by following the considerations. The first system is Parallel-Topics, a visual analytics system supporting analysis of large collections of documents by extracting semantically meaningful topics. Based on lessons learned from Parallel-Topics, this dissertation further presents a general visual text analysis framework, I-Si, to present meaningful topical summaries and temporal patterns, with the capability to handle large-scale textual information. Both systems have been evaluated by expert users and deemed successful in addressing domain analysis needs.;The second contribution lies in preserving domain users' analysis process while using interactive visualizations. Our research suggests the preservation could serve multiple purposes. On the one hand, it could further improve the current system. On the other hand, users often need help in recalling and revisiting their complex and sometimes iterative analysis process with an interactive visualization system. This dissertation introduces multiple types of evidences available for capturing a user's analysis process within an interactive visualization and analyzes cost/benefit ratios of the capturing methods. It concludes that tracking interaction sequences is the most un-intrusive and feasible way to capture part of a user's analysis process. To validate this claim, a user study is presented to theoretically analyze the relationship between interactions and problem-solving processes. The results indicate that constraining the way a user interacts with a mathematical puzzle does have an effect on the problemsolving process. As later evidenced in an evaluative study, a fair amount of high-level analysis can be recovered through merely analyzing interaction logs.