Using Global t-SNE to Preserve Intercluster Data Structure.

Area of Science:

• Data Science
• Computational Biology
• Bioinformatics

Background:

• t-distributed stochastic neighbor embedding (t-SNE) is a key technique for data visualization and clustering.
• t-SNE minimizes local data point distances but discards global structure information.
• Existing methods struggle to preserve both local and global data relationships.

Purpose of the Study:

• To develop a novel method, global t-SNE (g-SNE), that preserves global data structure alongside local clustering.
• To evaluate g-SNE's effectiveness in uncovering large-scale data organization.
• To compare g-SNE against established methods like t-SNE and UMAP.

Main Methods:

• Augmenting the t-SNE cost function with a global cost component.
• Applying g-SNE to synthetic, flower shape, and human brain cell datasets.
• Utilizing topological analysis to assess data distribution across scales.

Main Results:

• g-SNE successfully preserves significant global intercluster data structure.
• g-SNE demonstrates superior performance over t-SNE and UMAP in maintaining global topology.
• Meaningful global structures were identified in both plant and human brain datasets.
• Differences in data distribution across scales were observed between human brain datasets.

Conclusions:

• g-SNE effectively clusters data while preserving global structure, offering a more comprehensive view.
• The method reveals previously unidentified aspects of data organization across multiple scales.
• g-SNE provides a valuable tool for exploring complex biological and other high-dimensional datasets.