ColorPCA: Scalable Colored Dimensionality Reduction for Unlabeled High-Dimensional Data

Mapping labeled high-dimensional data to colors based on class labels in low-dimensional projections is effective for enhancing pattern recognition and reducing misinterpretation of clusters. However, automatic coloring of unlabeled high-dimensional data remains a challenge to reveal unknown patterns or class structures in the data. To address this, we propose ColorPCA, a scalable method that improves existing dimensionality reduction-based automatic coloring by integrating Principal Component Analysis with alpha compositing. Rather than mapping reduced dimensions to color coordinates, ColorPCA encodes data directly into color space to improve pattern discovery in unlabeled datasets. We implemented ColorPCA in a web-based visual analytics system for interactive exploration and evaluated it through three case studies using benchmark, simulated, and real-world climate datasets. Additionally, we conducted three user studies, two with generic users and one with climate domain experts. Comparisons with two state-of-the-art coloring methods based on PCA and t-SNE demonstrate that ColorPCA improves visual separability and facilitates deeper insight extraction in high-dimensional data visualization.