On Sketch-Based Selections From Scatterplots Using KDE, Compared to Mahalanobis and CNN Brushing

Area of Science:

• Computer Science
• Data Visualization
• Human-Computer Interaction

Background:

• Accurate data subset selection is vital for effective visual data exploration.
• Sketch-based methods, including kernel density estimation, offer promising solutions for interactive brushing.

Purpose of the Study:

• To introduce and detail a kernel density estimation-based brushing technique for scatterplots.
• To compare this technique with Mahalanobis brushing and Convolutional Neural Network (CNN) brushing.
• To evaluate the accuracy, efficiency, generality, and interpretability of empirical versus deep learning (DL) implicit modeling.

Main Methods:

• Developed a kernel density estimation (KDE) method for click-and-drag data subset selection in scatterplots.
• Conducted two user studies to compare KDE brushing with Mahalanobis brushing and CNN brushing.
• Performed a quantitative three-fold comparison and analyzed success/failure cases for each technique.

Main Results:

• The KDE brushing technique provides an effective method for data subset selection.
• Comparative analysis revealed differences in accuracy, efficiency, generality, and interpretability between empirical (KDE, Mahalanobis) and implicit (CNN) modeling approaches.
• Detailed prevalence of success and failure scenarios for each brushing method was identified.

Conclusions:

• Kernel density estimation presents a viable and efficient approach for interactive data brushing in scatterplots.
• The study offers insights into the trade-offs between different brushing techniques and modeling paradigms (empirical vs. implicit DL).
• Findings contribute to understanding the practical performance and user experience of various visual data exploration tools.