Supporting Analysis of Dimensionality Reduction Results with Contrastive Learning

Area of Science:

• Data Science
• Computer Science
• Statistics

Background:

• Dimensionality reduction (DR) is crucial for high-dimensional data analysis and visualization.
• Interpreting DR results requires understanding cluster characteristics, a current challenge.
• Existing methods for identifying clusters lack effective ways to characterize them.

Purpose of the Study:

• To present a visual analytics method for highlighting essential features of clusters in DR results.
• To address the challenge of understanding cluster characteristics in high-dimensional data.
• To introduce an enhanced usage of contrastive principal component analysis (cPCA) for feature contribution analysis.

Main Methods:

• Developed ccPCA (contrasting clusters in PCA) to calculate feature contributions between clusters.
• Integrated ccPCA into an interactive visual analytics system.
• Utilized scalable visualization for displaying cluster feature contributions.

Main Results:

• ccPCA effectively identifies and highlights features that characterize specific clusters.
• The interactive system provides a scalable visualization of feature contributions.
• Case studies on publicly available datasets demonstrate the method's effectiveness.

Conclusions:

• The ccPCA method enhances the interpretability of dimensionality reduction results.
• This visual analytics approach aids in gaining useful insights from complex, high-dimensional data.
• The developed system offers a practical solution for understanding cluster characteristics.