When visualizing large datasets, we inevitably run into a physical constraint: the digital display has a fixed number of pixels. As noted by Soon-Tee Teoh et al. in Visual-Based Anomaly Detection for BGP Origin AS Change (OASC) Events, mapping a massive data space to a fixed-resolution grid causes visual collisions where multiple distinct entities overlap on a single pixel.
This phenomenon, known as overplotting or visual congestion, creates what Diana Domańska et al. call a "congestion trap" in The rainfall plot: its motivation, characteristics and pitfalls. At high densities, a single pixel can represent thousands of overlapping events, causing the visual signal to plateau and masking the true intensity of the data. Whether exploring large machine learning embeddings (Fred Hohman et al. in WizMap: Scalable Interactive Visualization for Exploring Large Machine Learning Embeddings) or dense geospatial networks (Sujoy Bhore et al. in Worbel: Aggregating Point Labels into Word Clouds), this visual saturation hides internal distribution patterns and obscures critical outliers.
The Limits of Naive Solutions
A common instinct to fix overplotting is to use transparency (alpha-blending) or data filtering. However, these approaches introduce severe trade-offs:
- Alpha-Blending Artifacts: As Milos Krstajic et al. demonstrate in CloudLines: Compact Display of Event Episodes in Multiple Time-Series, transparency causes unpredictable opacity stacking when multiple markers coincide at the same coordinate. This creates perceptual artifacts that mislead users about the true underlying density.
- Information Loss: Filtering or sampling simply discards data. As Jinwook Seo et al. point out in Guest Editors' Introduction: Special Section on IEEE PacificVis 2019, traditional filtering loses critical information and outliers, driving the need for techniques that maintain data integrity while improving legibility.
- Interaction Paralysis: When points are clustered too closely, users face severe selection difficulties, as documented in PromptMap, making interactive exploration nearly impossible in crowded regions.
To overcome these limitations, researchers have developed two primary paradigms: Continuous Density Mapping and Spatial Relaxation.
Continuous Density Mapping and Topological Shading
One way to bypass the pixel-saturation limit is to stop drawing individual markers altogether and instead visualize the concentration of data points as a continuous field. As shown in CellProfiler Analyst: data exploration and analysis software for complex image-based screens, density-based visualization prevents visual saturation by encoding point concentration rather than individual markers, allowing researchers to identify clusters within millions of observations that would otherwise appear as a solid mass.
However, standard density plots or heatmaps introduce their own problem: they often appear flat, failing to highlight subtle topological features like ridges or valleys in medium-density regions.
[Raw Scatter Plot] ───► [Standard Density Plot] ───► [Illuminated Density Plot]
(Severe Overlap) (Flat, Saturation) (Shading Reveals Ridges)
To resolve this, Xin Chen et al. introduced Visualization-driven Illuminated Density Plots in Visualization-Driven Illumination for Density Plots. By combining a structure-aware shading model with a luminance-only color blending technique, they treat the density map like a 3D terrain. This allows users to see fine-grained topological details while still accurately reading density values from a color legend.
Similarly, Han Bao et al. proposed Bi-Scale Density Plots in Bi-Scale density-plot enhancement based on variance-aware filter, using a variance-aware filter to decompose density maps into base and detail layers. This separation allows users to see global trends and subtle local structures simultaneously without visual clutter or color saturation.
Spatial Relaxation and Density-Aware Displacement
If maintaining atomic access to individual data points is required, researchers turn to spatial relaxation—slightly adjusting point positions to eliminate overlap while preserving the global distribution. Jinwook Seo et al. describe this in Guest Editors' Introduction: Special Section on IEEE PacificVis 2019 as relaxing point layouts into space-filling maps, which preserves the visibility of all data points and outliers without discarding data.
The challenge lies in balancing clarity with spatial fidelity. As Thomas Low et al. note in The ISMIR Explorer - A Visual Interface for Exploring 20 Years of ISMIR Publications, snapping items to a rigid grid resolves clutter but distorts the precise spatial distances that represent semantic similarity.
To optimize this trade-off, several algorithmic approaches have emerged:
- Discrete Grid Optimization: The framework proposed in Optimal Neighborhood Preserving Visualization by Maximum Satisfiability uses Maximum Satisfiability (MaxSAT) to guarantee a globally optimal layout on a discrete grid, ensuring every data point remains distinct and accessible while respecting neighborhood constraints.
- Continuous Space Packing: Ruizhi Shi developed a dual space coupling model called PolarPacking in Dual Space Coupling Model Guided Overlap-Free Scatterplot. This approach generates overlap-free scatterplots that maintain high fidelity to the original data's density and shape, even when multiple data points share identical coordinates.
- Interactive Expansion: For targeted exploration, Zhibin Niu et al. introduced a radial expansion interaction in iConViz: Interactive Visual Exploration of the Default Contagion Risk of Networked-Guarantee Loans. When multiple data points share identical attributes, this mechanism spreads overlapping elements into a flower-like arrangement, ensuring individual entities remain accessible in high-density regions.
Managing Occlusion Across Dimensions
Moving from 2D to 3D is a tempting way to find more "room" for dense data, but it introduces severe visual occlusion and perspective distortion, as argued by Shaolun Ruan et al. in VENUS: A Geometrical Representation for Quantum State Visualization. In 3D cityscape metaphors or hierarchical graphs, foreground structures frequently hide critical background data points (Jiabin Zheng et al. in To Enlighten Hidden Facts in The Code: A Review of Software Visualization Metaphors; Yang Zhang et al. in Analyzing and Visualizing Scalar Fields on Graphs).
To combat this in spatiotemporal data, Jiabao Huang et al. developed VolumeSTCube in Volume-Based Space-Time Cube for Large-Scale Continuous Spatial Time Series. Instead of using discrete markers that clutter and hide behind one another, they transform sensor data into a continuous 3D volume. By applying volume and surface rendering techniques, they solve visual occlusion and depth ambiguity, allowing users to track complex phenomena across millions of data points smoothly.
Key Foundational Work
Visual Data Mining of Large Spatial Data Sets (2003) — contributed 4 concept elements to this direction.
The Gridfit algorithm: an efficient and effective approach to visualizing large amounts of spatial data (1998) — contributed 3 concept elements to this direction.
Visual-Based Anomaly Detection for BGP Origin AS Change (OASC) Events (2003) — contributed 1 concept elements to this direction.