VisExPreS: A Visual Interactive Toolkit for User-Driven Evaluations of Embeddings.

Area of Science:

• Data Science
• Computer Vision
• Machine Learning

Background:

• Dimensionality reduction is crucial for big-data analytics but often acts as a black box, making interpretation and evaluation challenging.
• Existing methods for analyzing low-dimensional embeddings include quantitative approaches lacking user control and visual techniques heavily reliant on user expertise.

Purpose of the Study:

• To present VisExPreS, a novel visual interactive toolkit designed for user-driven assessment of low-dimensional embeddings.
• To bridge the gap between purely quantitative and purely visual methods for embedding analysis.

Main Methods:

• VisExPreS employs three new techniques: PG-LAPS, PG-GAPS, and RepSubset.
• These methods generate interpretable explanations of local and global structures within embeddings.
• The system proactively guides users through the analysis process.

Main Results:

• VisExPreS facilitates a more controlled and interpretable assessment of embeddings.
• The toolkit effectively aids in interpreting, analyzing, and evaluating embeddings from various dimensionality reduction algorithms.
• User studies confirmed the utility and effectiveness of VisExPreS.

Conclusions:

• VisExPreS enhances the analysis of low-dimensional embeddings by integrating user guidance with interpretable structural explanations.
• The toolkit offers a valuable solution for overcoming the interpretability and evaluation challenges in dimensionality reduction.
• VisExPreS empowers users to conduct more informed and objective assessments of embedding quality.