Structure preserving t-SNE of matrix framed data.

Area of Science:

• Data Visualization
• Machine Learning
• Dimensionality Reduction

Background:

• Matrix-framed data, with elements indexed by rows and columns, is common across scientific fields.
• Current visualization techniques often overlook the inherent 2D structure of matrix-framed data.
• Representing diverse data types (scalars, vectors, time series, matrices, arrays) within this structure poses challenges.

Purpose of the Study:

• To introduce a novel visualization method, Matrix t-SNE, specifically designed for matrix-framed data.
• To effectively embed matrix elements into a low-dimensional Euclidean space.
• To preserve both row-wise and column-wise group structures within the embedded data.

Main Methods:

• Extension of the classical t-Distributed Stochastic Neighbor Embedding (t-SNE) algorithm.
• Development of a detailed algorithmic framework for embedding matrix-framed data.
• Application and evaluation on three diverse real-world datasets: exergame, gene expression, and temperature.

Main Results:

• Matrix t-SNE demonstrates superior performance in separating data elements based on latent row and column structures.
• The method effectively embeds matrix elements while preserving their group affiliations.
• Comparative analysis shows Matrix t-SNE's advantage over classical t-SNE in capturing matrix-framed data characteristics.

Conclusions:

• Matrix t-SNE offers a significant advancement for visualizing matrix-framed data.
• The method provides a robust approach to dimensionality reduction while respecting data's 2D organization.
• This technique enhances the interpretability of complex datasets with inherent row and column relationships.