A General Framework for Comparing Embedding Visualizations Across Class-Label Hierarchies

Area of Science:

• Data visualization
• Machine learning
• Computational biology

Background:

• Embedding visualization aids high-dimensional data interpretation.
• Current comparison methods require direct point correspondences, limiting broader applications.
• Existing techniques fail to capture complex relationships between point groups.

Purpose of the Study:

• To develop a general framework for comparing embedding visualizations using shared class labels.
• To characterize intra- and inter-class relationships by partitioning points into confusion, neighborhood, and relative size regions.
• To enable meaningful comparisons across different datasets and label hierarchies.

Main Methods:

• Developed a framework partitioning points into class-based regions (confusion, neighborhood, relative size).
• Utilized perceptual neighborhood graphs to define these regions.
• Introduced quantitative metrics to characterize intra- and inter-class relationships.

Main Results:

• Demonstrated framework generality in machine learning and single-cell biology use cases.
• Highlighted metrics' ability to provide insightful comparisons across label hierarchies.
• Evaluation study showed increased participant confidence and structured comparisons.

Conclusions:

• The proposed framework offers a robust method for comparing embedding visualizations without point correspondences.
• The class-label-based approach enhances understanding of complex data relationships.
• This method improves decision-making and interpretation in diverse scientific domains.