Supervised t-distributed stochastic neighbor embedding for data visualization and classification.

Area of Science:

• Machine Learning
• Data Science
• Computational Statistics

Background:

• High-dimensional data presents challenges for traditional dimension reduction techniques.
• Existing methods may struggle with prediction and visualization in ultra-high dimensional settings (p > n).

Purpose of the Study:

• To propose a novel supervised dimension reduction method, St-SNE.
• To enhance both prediction and visualization capabilities for high-dimensional data.
• To introduce a criterion for automatic selection of reduced dimension size.

Main Methods:

• Supervised t-distributed stochastic neighbor embedding (St-SNE) preserves similarities in feature and outcome spaces.
• The method handles high-dimensional data, including settings where p > n.
• A penalized Kullback-Leibler divergence criterion is proposed for selecting the reduced dimension k.

Main Results:

• St-SNE demonstrates superior prediction performance in ultra-high dimensional settings (p > n).
• It shows competitive prediction performance in p <= n settings.
• St-SNE proves to be a capable visualization tool, capturing within-cluster variations.

Conclusions:

• St-SNE offers a powerful approach for supervised dimension reduction.
• The method effectively balances prediction and visualization for diverse datasets.
• Automatic dimension selection enhances the practical applicability of St-SNE.