Mesoscopic structure graphs for interpreting uncertainty in non-linear embeddings.

Area of Science:

• Data Visualization
• Machine Learning
• Computational Biology

Background:

• Non-linear dimensionality reduction (NLDR) methods like t-SNE and UMAP are crucial for visualizing high-dimensional data.
• These methods can introduce distortion errors, leading to inaccurate interpretations of data structures.
• Existing visualization techniques struggle with large datasets and unsupervised analysis.

Purpose of the Study:

• To address the limitations of current NLDR visualization methods.
• To propose a novel visualization technique, ManiGraph, for improved neighborhood fidelity and interpretation of DR results.
• To tackle overplotting issues in large-scale datasets and support unsupervised analysis.

Main Methods:

• Conducted a survey of existing layout enrichment visualizations for DR.
• Developed ManiGraph, a node-link visualization technique.
• Constructed dynamic mesoscopic structure graphs and measured region-adapted trustworthiness to assess neighborhood fidelity.

Main Results:

• ManiGraph effectively reduces distortion errors in dimensionality reduction visualizations.
• The technique successfully addresses overplotting in large datasets.
• Demonstrated ManiGraph's utility across diverse applications, including machine learning, single-cell RNA sequencing, and histopathology image analysis.

Conclusions:

• ManiGraph offers a more reliable approach to interpreting complex high-dimensional data visualized with DR techniques.
• The method enhances the fidelity of neighborhood relationships between high- and low-dimensional spaces.
• ManiGraph provides a robust solution for data exploration in unsupervised and large-scale scenarios.