Idea

Compute pairwise similarities in high-dim (Gaussian) + low-dim (t-distribution). Minimize KL divergence between distributions.

Advertisement

Perplexity

Effective neighborhood size. Typical 30. Different perplexities reveal different structure — try multiple.

Advertisement

Caveats

Cluster sizes not meaningful. Distances between clusters not meaningful. Random init → different runs different.

UMAP

Alternative: faster + preserves more global structure. Now often preferred over t-SNE. Same 2D visualization purpose.