Cheatsheet: Seeing high-dimensional data: t-SNE

What t-SNE is

Item	Detail
Purpose	visualization only (not preprocessing)
Output	a 2D layout where similar high-D points appear near each other
How it works	match 2D pairwise similarities to high-D pairwise similarities, iteratively
Strong example	MNIST: PCA smears digits, t-SNE produces 10 separated blobs

What it preserves (and does not)

Aspect	Preserved?
Local structure (who is near whom inside a cluster)	yes
Global structure (distance/arrangement between clusters)	no
Cluster size on the page	no (artifact of layout)
Layout across runs	no (varies with seed)

Three misreadings to avoid

Misreading	What’s actually true
”Clusters drawn close together are more related”	between-cluster distance is meaningless
”Big cluster = more variation”	cluster size on the page is meaningless
”Same data always gives the same picture”	different seeds give different layouts

Perplexity

Setting	Effect
Too low (e.g., 5 on a non-tiny dataset)	many tiny fragmented clusters
Too high (e.g., 100+)	everything blurs into one blob
Typical range	5 to 50
Best practice	try several; trust clusters that appear stably

t-SNE vs PCA

	PCA	t-SNE
Goal	compression (keep variance)	visualization (preserve local neighborhoods)
Linearity	linear	nonlinear
Use for modeling?	yes	no
Preserves global structure	yes	no
Preserves local clusters	weakly	strongly

Cousin: UMAP

Note
Often faster than t-SNE	preserves more global structure than t-SNE
When you have a choice	try both