Search

Visually exploring data is a powerful approach to discover, understand, and interpret patterns that might otherwise go unnoticed. However, as datasets increase in size and complexity, visual pattern exploration can become challenging. For instance, patterns of interest can appear frequently while being distributed sparsely. Related pattern types can differ in size by several orders of magnitude. And patterns can be detected with high uncertainty. In the context of this dissertation, a pattern can be any reoccurring visual artifact, like a peak in a bar chart, that acts as a proxy for a relevant event or data property. In general, while scalability has been a core research topic in visualization, current tools have limited support for exploring large numbers of reoccurring patterns at and across scales. This dissertation studies scalable visual pattern exploration through a series of new visualization systems and interaction techniques for three essential tasks: browsing, comparing, and finding patterns. By treating pattern instances as first-class objects during the exploration and guiding this process with human-centered machine learning, these visualization tools enable exploring large pattern spaces. Building upon preliminary work on a web-based platform for multiscale visualization of large quantitative datasets, the first technique that this dissertation introduces is called Scalable Insets. Scalable Insets provides pattern-driven guidance in navigating multiscale visualization to enable finding regions of interest more efficiently. For comparing local patterns, this dissertation presents HiPiler—a system for visually decomposing multiscale matrix visualizations into a series of interactive small multiples. This enables effective clustering and quality control through grouping and aggregation. Based on the findings from the first two systems, this dissertation describes a generalized design space for interactive visual piling. Piling is an exploration approach to interactively arrange, group, and visually aggregate thousands of pattern instances in a small multiples setup. And finally, this dissertation demonstrates how unsupervised deep representation learning can be combined with interactive visual machine learning to enhance our ability to find complex or not well-defined patterns in large sequential datasets. In combining pattern-driven exploration with machine learning-powered guidance, these systems ensure that human-in-the-loop data analysis remains feasible with increasingly large and complex datasets.