HUMAP: Hierarchical Uniform Manifold Approximation and Projection.

Area of Science:

• Data Science
• Information Visualization
• Machine Learning

Background:

• Dimensionality reduction (DR) is crucial for pattern identification in high-dimensional data.
• Hierarchical DR techniques are ideal for datasets with multiple granularities, offering a 'major structures first, details on demand' approach.
• Existing hierarchical DR methods may struggle with preserving both local and global data structures simultaneously.

Purpose of the Study:

• To introduce HUMAP, a novel hierarchical dimensionality reduction technique.
• To develop a flexible DR method that preserves both local and global data structures.
• To maintain the 'mental map' during hierarchical data exploration.

Main Methods:

• Development of the HUMAP hierarchical dimensionality reduction technique.
• Empirical evaluation comparing HUMAP against existing hierarchical DR approaches.
• Application of HUMAP in a case study for dataset labeling.

Main Results:

• HUMAP demonstrates superior performance compared to current hierarchical DR techniques.
• The technique effectively balances the preservation of local and global data structures.
• Successful application of HUMAP for efficient dataset labeling.

Conclusions:

• HUMAP offers a flexible and effective hierarchical dimensionality reduction solution.
• The method enhances understanding of complex, high-dimensional datasets.
• HUMAP shows promise for practical applications like data labeling.