Self-Organizing Nebulous Growths for Robust and Incremental Data Visualization.

Area of Science:

• Data Science
• Machine Learning
• Computer Vision

Background:

• Nonparametric dimensionality reduction methods like t-SNE and UMAP excel at visualizing static datasets.
• These methods lack the capability to incrementally update visualizations with new data points.
• Existing techniques struggle to integrate new data without compromising the integrity of existing visualizations.

Purpose of the Study:

• Introduce Self-Organizing Nebulous Growths (SONG), a novel parametric nonlinear dimensionality reduction technique.
• Enable incremental data visualization, allowing for the seamless addition of new data points.
• Demonstrate SONG's ability to handle both similar and heterogeneous data increments while preserving visualization structure.

Main Methods:

• Developed SONG, a parametric nonlinear dimensionality reduction algorithm.
• Tested SONG on diverse real-world and simulated datasets.
• Compared SONG's performance against Parametric t-SNE, t-SNE, and UMAP using cluster quality metrics like adjusted mutual information.

Main Results:

• SONG significantly outperforms Parametric t-SNE, t-SNE, and UMAP in incremental data visualization tasks.
• Achieved substantial improvements in cluster quality for heterogeneous increments (up to 49.73% on MNIST).
• Demonstrated superior or comparable performance to UMAP and t-SNE even when data is presented all at once, alongside enhanced noise tolerance.

Conclusions:

• SONG offers a robust solution for dynamic and evolving datasets requiring incremental visualization.
• The parametric nature and algorithmic design of SONG provide superior performance and noise resilience.
• SONG represents a significant advancement in dimensionality reduction for real-time and high-variance data analysis.