Conditional t-SNE: more informative t-SNE embeddings.

Area of Science:

• Data Science
• Machine Learning
• Computational Statistics

Background:

• Dimensionality reduction techniques like t-distributed stochastic neighbor embedding (t-SNE) are crucial for visualizing high-dimensional data.
• Existing methods face limitations: inability to capture all information in a single 2D embedding and redundancy for informed users, hindering novel insights.

Purpose of the Study:

• To address the limitations of current dimensionality reduction methods.
• To introduce a generalized t-SNE approach that incorporates prior information to yield more informative embeddings.

Main Methods:

• Introduction of conditional t-SNE (ct-SNE), a generalization of t-SNE.
• Development of a conditioned t-SNE objective function that discounts prior information (e.g., labels).
• Efficient optimization strategies for the ct-SNE objective were proposed and analyzed.

Main Results:

• ct-SNE effectively discounts prior information, enabling the capture of complementary structures.
• Empirical results on synthetic and real-world datasets demonstrate ct-SNE's scalability and effectiveness.
• The method successfully generated more informative and relevant embeddings compared to standard t-SNE.

Conclusions:

• Conditional t-SNE offers a powerful approach to extract previously inaccessible information from high-dimensional data.
• The technique facilitates the discovery of novel insights by revealing structures missed by conventional embedding methods.
• ct-SNE represents a significant advancement in manifold learning for data exploration and analysis.