Parallel Clusters: Visual Comparison of Embeddings Based on Multi-Scale Neighborhood Analysis.

Area of Science:

• Data Visualization
• Machine Learning Interpretability
• Visual Analytics

Background:

• Comparing high-dimensional data embeddings visually is challenging due to difficulties in aligning structures across different views.
• Existing methods often struggle to provide intuitive comparisons of complex embedding relationships.

Purpose of the Study:

• To introduce Parallel Clusters (paraClus), a visual analytics system designed for multi-scale exploration and comparison of embedding structures.
• To address the challenge of visually comparing embeddings by enabling meaningful comparisons through multiple data hierarchies.

Main Methods:

• paraClus constructs data hierarchies from multiple perspectives, facilitating comparison.
• It employs cluster formations across embeddings and attributes with an adaptive thresholding mechanism for neighborhood definition.
• A parallel axes design aligns clusters for easy comparison of neighboring structures and relationships across embeddings.

Main Results:

• The system allows users to explore embedding structures at varying scales by dynamically adjusting the threshold.
• Interactive subdivision mechanisms enable refinement of clusters based on inter-cluster connections, revealing structural dependencies.
• paraClus integrates labels and scalar attributes for unified analysis of diverse embedding types, including multi-attribute and time-varying data.

Conclusions:

• paraClus effectively enhances the understanding of embedding relationships for neural network interpretability, structural analysis, and data exploration.
• Expert assessments and case studies validate the system's utility in analyzing complex, multi-attribute, and time-varying embeddings.