Evaluation of Sampling Methods for Scatterplots.

Area of Science:

• Data Visualization
• Computer Graphics
• Human-Computer Interaction

Background:

• Large-scale scatterplots present challenges for visual analysis due to data volume.
• Effective sampling is necessary to create representative and informative smaller scatterplots.
• Understanding how sampling strategies impact visual perception is key for abstraction.

Purpose of the Study:

• To evaluate the effectiveness of different sampling strategies for multi-class scatterplots.
• To determine which sampling methods best preserve density, outliers, and overall shape.
• To provide guidance on selecting appropriate sampling techniques for scatterplot abstraction.

Main Methods:

• Literature review to identify seven typical sampling strategies.
• Selection of eight representative multi-class scatterplot datasets.
• Design and execution of four experiments focusing on region density, class density, outlier preservation, and shape maintenance.

Main Results:

• Random sampling is optimal for preserving region density.
• Blue noise and random sampling show comparable performance in preserving class density.
• Outlier-biased density-based, recursive subdivision-based, and blue noise sampling are superior for outlier preservation.
• Blue noise sampling demonstrates the best performance in maintaining the overall scatterplot shape.

Conclusions:

• The choice of sampling strategy significantly impacts the fidelity of scatterplot abstractions.
• Blue noise sampling emerges as a robust method for preserving key visual characteristics like shape and outliers.
• Different sampling methods are suited for different abstraction goals, such as preserving overall density versus specific features.