Remove, Replace, Refine: A Computational Approach to Improving Color Palettes for Data Visualization

Abstract

Color plays a critical role in how we perceive and interpret data visualizations such as graphs, maps, and charts. For example, a well-curated color palette can make charts both visually appealing and easy to understand. But, what makes a palette more aesthetically preferable? This project investigates the perceptual features that predict people’s color palette preferences in data visualizations using behavioral experiments and computational modeling. Participants compared pairs of three-color palettes displayed in data visualizations and rated their preference on a continuous scale. A regression model built from lightness and hue features explained about 70% of variance in preference ratings. Results suggested that people preferred palettes where lightness was distributed more uniformly across colors and leaned toward cooler hues like cyan-green, while shifts toward warmer hues like orange were less preferred. To test generalization, we used the model to select palette modifications predicted to be preferred over their originals. New participants preferred the model-selected modifications in 19 of 22 cases, demonstrating that the model generalized to new palettes judged by new participants. These findings both offer insight into how people evaluate color combinations in data visualizations and can be used to improve color palette recommendation tools used for data visualization (i.e., Colorgorical).