Data Visualization with Structural Control of Global Cohort and Local Data Neighborhoods

Area of Science:

• Computer Science
• Data Science
• Machine Learning

Background:

• Effective data visualization aims to reveal local and global data structures.
• Existing methods struggle to balance local neighborhood preservation with cohort separation.

Purpose of the Study:

• To develop a novel visualization framework enhancing both local sample structure and global cohort positioning.
• To provide better control over the arrangement and proximity of data cohorts in low-dimensional representations.

Main Methods:

• Incorporation of cohort positioning and discriminative constraints using computed cohort prototypes.
• Development of embedding and projection algorithms optimized via matrix manifold procedures.
• Proposal of a matrix decomposition model for accelerated computation in large-scale applications.

Main Results:

• Demonstrated improved capabilities of the new methods with state-of-the-art dimensionality reduction algorithms.
• Qualitative and quantitative comparisons on synthetic and real-world text/image data show notable improvements.
• Effective control over cohort arrangements and proximities in generated visualizations.

Conclusions:

• The proposed framework offers superior data visualization by integrating cohort-specific information.
• The methods provide enhanced insights into both local sample relationships and global cohort structures.
• This approach advances dimensionality reduction techniques for complex datasets.