| Visualization has emerged as an effective means to quickly obtain insight from raw data. While simple computer programs can generate simple visualizations, and while there has been constant progress in sophisticated algorithms and techniques for generating insightful pictorial descriptions of complex data, the process of building visualizations remains a major bottleneck in data exploration.;In this thesis, we present the main design and implementation aspects of VisTrails, a system designed around the idea of transparently capturing the exploration process that leads to a particular visualization. In particular, VisTrails explores the idea of provenance management in visualization systems: keeping extensive metadata about how the visualizations were created and how they relate to one another.;This thesis presents the provenance data model in VisTrails, which can be easily adopted by existing visualization systems and libraries. This lightweight model entirely captures the exploration process of the user, and it can be seen as an electronic analogue of the scientific notebook. The provenance metadata collected during the creation of pipelines can be reused to suggest similar content in related visualizations and guide semi-automated changes. This thesis presents the idea of building visualizations by analogy in a system that allows users to change many visualizations at once, without requiring them to interact with the visualization specifications.;It then proposes techniques to help users construct pipelines by consensus, automatically suggesting completions based on a database of previously created pipelines. By presenting these predictions in a carefully designed interface, users can create visualizations and other data products more efficiently because they can augment their normal work patterns with the suggested completions.;VisTrails leverages the workflow specifications to identify and avoid redundant operations. This optimization is especially useful while exploring multiple visualizations. When variations of the same pipeline need to be executed, substantial speedups can be obtained by caching the results of overlapping subsequences of the pipelines. We present the design decisions behind the execution engine, and how it easily supports the execution of arbitrary third-party modules. These specifications also facilitate the reproduction of previous results. We will present a description of an infrastructure that makes the workflows a complete description of the computational processes, including information necessary to identify and install necessary system libraries. In an environment where effective visualization and data analysis tasks combine many different software packages, this infrastructure can mean the difference between being able to replicate published results and getting lost in a sea of software dependencies and missing libraries.;The thesis concludes with a discussion of the system architecture, design decisions and learned lessons in VisTrails. This discussion is meant to clarify the issues present in creating a system based around a provenance tracking engine, and should help implementors decide how to best incorporate these notions into their own systems. |
