t-SNE
t-SNE maps high-dimensional data to low dimensions (2-3D) while preserving local neighborhood structure; it is the first choice for visualizing high-dimensional data.
Purpose
To visualize high-dimensional data via nonlinear dimensionality reduction that preserves local structure.
Mechanism
Similarities are modeled with Gaussians in high dimensions and with the t-distribution in low dimensions (heavy tails avoid crowding). It minimizes KL divergence. Complexity is O(n^2); large datasets use the Barnes-Hut approximation at O(n log n). Use it for visualization only, not as general features.
Use cases
- High-dimensional data visualization
- Manifold learning
- Data exploration