Joint t-SNE for Comparable Projections of Multiple High-Dimensional Datasets

Area of Science:

• Computational science
• Data visualization
• Machine learning

Background:

• t-Stochastic Neighbor Embedding (t-SNE) is widely used for visualizing high-dimensional data.
• Independent t-SNE projections of multiple datasets yield misaligned layouts, hindering comparisons.
• Existing methods struggle to maintain consistent visualizations across datasets, especially over time.

Purpose of the Study:

• To introduce Joint t-Stochastic Neighbor Embedding (Joint t-SNE) for generating comparable projections of multiple high-dimensional datasets.
• To enable direct visual comparisons between different datasets or time points.
• To address the limitations of standard t-SNE in comparative analysis.

Main Methods:

• Introduced edge similarity using Graphlet Frequency Distribution (GFD) to capture similarities between adjacent data frames.
• Integrated a novel loss term, 'vector constraints,' into the t-SNE loss function.
• Preserved vectors between projected points across multiple projections to serve as visual landmarks.

Main Results:

• Joint t-SNE demonstrated superior performance over existing techniques like Dynamic t-SNE on synthetic datasets.
• Outperformed in metrics including local coherence error, Kullback-Leibler divergence, and neighborhood preservation.
• Successfully visualized and compared neural network layer activations in a real-world use case.

Conclusions:

• Joint t-SNE effectively generates comparable projections for multiple high-dimensional datasets.
• The technique overcomes misalignment issues inherent in independent t-SNE applications.
• Enables robust comparative visualization for time-varying data and multi-set analyses.