Introduction
Visualizing large-scale line data—such as stock prices, temperature changes, or GPS trajectories—often results in a cluttered "hairball" that is difficult to interpret. While researchers frequently use density plots (heatmaps for lines) to reduce this clutter, these plots introduce a new set of problems: visual ambiguities.
Traditional density plots aggregate data into a single color scale, which can trigger misleading Gestalt perceptions. This paper identifies three primary "visual confusers" that lead to incorrect data interpretation:
- Illusory patterns: High-density regions appear where no single coherent trend exists.
- Ambiguous continuation: At intersections, it is impossible to determine which line follows which path.
- Disconnected clusters: Related trends appear separate because they are not perfectly adjacent in space.
Methodology: Image-Space Colorization
To resolve these ambiguities, the authors propose a pipeline that shifts clustering from the data-space (clustering the lines themselves) to the image-space (clustering the pixels of the plot). This approach focuses on what the user actually sees rather than the underlying raw data, which is often noisy.
The Pipeline
- Feature Extraction: The canvas is divided into bins (pixels). For each bin, the algorithm records a "feature set" consisting of the IDs of every line passing through that pixel.
- Hierarchical Clustering: Bins are grouped together if they share a similar set of lines. The authors utilize the Overlap Coefficient instead of standard metrics like Jaccard similarity. This allows high-density "main roads" and lower-density "side streets" to be grouped into the same trend if one set is a subset of the other.
- Circular MDS Colorization: To assign colors, Circular Multidimensional Scaling (MDS) maps the high-dimensional clusters onto a 2D hue circle. This ensures that clusters with similar line compositions receive perceptually similar colors.
- Line Assignment: Finally, individual lines can be mapped back to these clusters, allowing users to filter raw data based on the discovered visual trends.
Key Results and Case Studies
A user study with 25 participants confirmed that traditional density plots are highly ambiguous; participants often disagreed on the number of trends present but remained "moderately to extremely certain" of their incorrect interpretations.
Beijing Taxi Trajectories
In a dataset of 10,357 taxis in Beijing, a traditional density plot (Figure b above) suggests a continuous circular route around the city center. However, the proposed colorization method (Figure c) debunks this illusion, revealing that taxis actually cluster in specific local sectors of the city. What appeared to be a single circular trend was actually the overlap of many short, localized trips.
Stock Market Data
When applied to NYSE stock data (2005–2017), the colorization revealed a "disconnected cluster" of stocks newly listed after 2010. This group appeared as a separate high-density trend that was completely invisible in standard monochromatic density plots.
Conclusion
By moving clustering into the image-space and applying perceptually uniform color mapping, this method effectively resolves the Gestalt-driven ambiguities inherent in traditional density plots. While limited by the number of distinguishable hues (roughly seven), the approach provides a significantly clearer view of overlapping trends in complex, noisy datasets.
Summary updated