Co-Adaptive Guidance in Mixed-Initiative Visual Analytics

Abstract

In a highly interconnected and optimized world, huge amounts of data can be generated every second. However, the pure collection of generated data typically does not provide increased value. Instead, the generation of this data is only meaningful if intelligent, tailored analysis can be employed to generate insights and derive decisions. In this direction, visual analytics (VA) is a research field that combines information visualization with data analysis. In particular, recent developments in machine learning (ML) and artificial intelligence (AI) techniques can be integrated into VA approaches to support users in their analysis. Typically, such support is provided in the form of guidance, where the VA system aims to help the user overcome potential knowledge gaps. The biggest challenge in guiding users is to correctly identify their intent and provide the correct help at the right time. This includes figuring out the specific analysis task users are performing, in which visual representation suggestions are most effective, and—more broadly—where their knowledge gaps are. In this dissertation, we focus on the topic of co-adaptation of both the user and the system in VA. Co-adaptation describes how, during the guidance process, both users and systems learn from each other and teach each other before, ultimately, converging to a common understanding of the task and the solution. We aim to answer the question of how such co-adaptive guidance systems can be achieved in practice. This dissertation can be split into three parts: First, we introduce the theory of co-adaptive guidance. Second, we bridge the gap between guidance theory and practical implementation in VA systems. Third, we present two applications showcasing different aspects of co-adaptation. In the first part, we contribute a process model of co-adaptation that highlights where and how adaptation can happen and argue for well-defined learning goals for adaptive systems that we derive from Bloom's taxonomy. This results in an intricate theoretical model. In the second part, we address the common disconnect between guidance research and what is practically implemented in systems. To overcome this issue, we present Lotse, a library for simplified guidance implementation built around guidance strategies. To determine which strategies can be employed and which interaction dynamics they cause, we also present the design and results of a Wizard of Oz study. In the final part, we employ the previous insights on co-adaptive guidance in two guidance applications for topic model refinement and argumentation annotation. The first application focuses on adapting in which contexts guidance is provided, while the second application adapts the content of provided suggestions. Our research provides a robust theoretical foundation for co-adaptive guidance in visual analytics and translates these concepts into practical tools and applications. By doing so, it bridges the gap between abstract theory and tangible implementation, offering a comprehensive solution that can be readily adopted and applied in various domains. Ultimately, this dissertation contributes to the ongoing evolution of visual analytics, paving the way for more intuitive, adaptive, and user-centric systems.