Shape-aware stochastic neighbor embedding for robust data visualisations.

Area of Science:

• Data visualization
• Machine learning
• Bioinformatics

Background:

• t-distributed Stochastic Neighbor Embedding (t-SNE) is widely used for high-dimensional data visualization, particularly in single-cell transcriptomics.
• t-SNE struggles with accurately representing hierarchical relationships and can create spurious patterns.
• Existing methods like UMAP and PHATE have limitations in addressing t-SNE's shortcomings.

Purpose of the Study:

• To generalize t-SNE using shape-aware graph distances to overcome its limitations.
• To develop a method that accurately visualizes hierarchical structures in high-dimensional data.
• To provide a robust alternative for dimensionality reduction and data visualization.

Main Methods:

• Generalization of t-SNE using shape-aware graph distances.
• Application of the proposed method to simulated and real-world datasets (single-cell transcriptomics, MNIST).
• Quantitative validation using established indices and comparison with t-SNE, UMAP, and PHATE.

Main Results:

• The proposed method significantly outperforms t-SNE, UMAP, and PHATE on imbalanced, nonlinear, continuous, and hierarchically structured data.
• Faithful low-dimensional embeddings were achieved on both simulated and real-world datasets.
• The method's single hyper-parameter can be automatically and optimally selected in a data-driven manner.

Conclusions:

• The shape-aware stochastic neighbor embedding method provides robust and accurate low-dimensional visualizations.
• This method effectively reveals key structures within high-dimensional data.
• It offers a significant improvement over existing dimensionality reduction techniques for complex datasets.