Automating Effective Categorical Palette Design: Empirically Grounded Tool for Color, Shape, and Redundant Encodings

Designing effective categorical palettes---sets of colors, shapes, or their combinations used to distinguish categories in visualizations — is essential for accurate data interpretation. However, this task remains challenging: perceptual limitations, increasing category numbers, and the diversity of real-world analytical goals make it difficult to ensure categories remain visually distinct. Existing design guidelines are largely heuristic, rarely grounded in empirical evidence across analytic tasks, encoding channels, or category scales. As a result, they provide limited insight into how palette choices affect human performance across different contexts. To address this gap, this dissertation argues that effective categorical palette design must be grounded in empirical models of human perception that account for variation across encoding channels, category counts, and analytical tasks. Rather than relying on general-purpose heuristics, it proposes a systematic, data-driven approach to understanding and optimizing palette effectiveness.This dissertation makes categorical palette design empirically grounded and actionable, developing perceptual models that explain human performance across encoding channels and translate directly into practical design tools. Across a series of studies, I examine how encoding channel, category count, and analytical task jointly shape the effectiveness of categorical palette design. I begin by investigating color as a categorical encoding channel. Through a crowdsourced experiment covering two to ten categories and ten widely-used color palettes, I show that both category count and palette choice significantly impact people's ability to compare category means, with performance degrading at different rates across palettes. Critically, these findings challenge dominant design heuristics: lightness magnitude, lightness variation, and overall perceptual distances, not hue contrast alone, predict robust palette performance across category counts. I extend this work by comparing five families of color design---categorical, sequential, diverging, and perceptually-uniform variants---showing that while hue-based categorical palettes remain most effective overall, lightness and perceptual uniformity provide meaningful additional benefits at larger category numbers. I then investigate shape as a categorical encoding channel through four experiments evaluating 39 shapes across mean estimation, expert preference, and correlation judgment tasks. Shape effectiveness cannot be predicted from broad perceptual categories such as filled versus unfilled, nor from designer intuition. Shape is complex that no simple feature set defines an effective shape palette. Instead, I construct a model based on pairwise perceptual relationships between shapes and embed it into Shape It Up, an interactive shape palette design tool that gives designers data-driven guidance for up to ten categories. Building on these single-channel findings, I examine whether redundant color–shape encodings improve task accuracy through four crowdsourced experiments. The result shows that redundancy is not uniformly beneficial: it overall improves accuracy for correlation tasks, but with the strongest gains for five to eight categories. I further show strong interaction effects between color and shape channels---naively combining top-performing color and shape palettes does not reliably produce top-performing redundant encodings, requiring careful cross-channel pairing. These findings are operationalized in CatPAW, a palette design tool that automates effective redundant palette generation by modeling these interactions. Finally, I unify these contributions in CatPAL, a task-aware categorical palette recommendation system. CatPAL estimates task-specific perceptual strengths for a set of colors and shapes using Bradley-Terry models, refined through adaptive sampling. A redundant gain metric captures cross-channel interactions across category counts and tasks. A predictive model trained on this multi-stage data scores candidate palettes by task, category count, and perceptual features derived from our previous works, like lightness variation, demonstrating that palette effectiveness is inherently task-dependent and that no single palette is optimal across contexts.Together, this work moves categorical palette design beyond heuristic convention toward empirically grounded, task-aware practice. By translating large-scale behavioral evidence into interactive recommendation tools, this dissertation provides both fundamental insights into human categorical perception and practical systems for improving real-world data visualization.