datarekha
Patterns June 10, 2026

Your t-SNE plot is lying to you (three ways)

t-SNE and UMAP reveal clusters PCA hides — but cluster sizes, the gaps between clusters, and even the shapes are often artifacts. How to read these plots without fooling yourself, and when to reach for UMAP instead.

6 min read · by datarekha · machine-learningtsneumapvisualization

PCA is linear, so curved structure flattens and overlapping clusters smear together. t-SNE and UMAP fix that — they’re nonlinear projections built to make similar high-dimensional points land near each other, revealing clusters PCA hides. They’re indispensable for inspecting embeddings. They’re also some of the most misread plots in all of ML.

Three things the plot is not telling you

The canonical warning is Distill’s interactive “How to Use t-SNE Effectively,” and its lessons are blunt. First, sizes: you cannot see the relative sizes of clusters in a t-SNE plot, because the algorithm evens out cluster sizes by adapting to local density. A big blob is not a big class. Second, distances: the distances between well-separated clusters in a t-SNE plot may mean nothing, so inter-cluster distance should not be interpreted. Two clusters far apart aren’t “more different.” Third, stability: t-SNE results depend heavily on perplexity, and seeing global geometry requires fine-tuning it rather than trusting a single value. Change the seed or the perplexity and the picture changes. Run it a few times before you believe any shape.

UMAP wins on scale — but isn’t magically more global

For large datasets, UMAP has become the default, and the speed gap is enormous. On identical hardware, benchmarks show UMAP projecting 100,000 points in about 2.4 minutes versus t-SNE’s 3.7 hours, and a million points in about 8.2 minutes versus roughly 14 days for t-SNE — a result of UMAP’s roughly O(n log n) complexity via approximate nearest neighbors and SGD versus t-SNE’s O(n²). But don’t over-trust UMAP’s “global structure” reputation: research finds UMAP does not preserve global structure any better than t-SNE when both use the same informative initialization.

These plots are powerful precisely because they exaggerate local neighborhood structure. That same exaggeration is what makes their sizes, gaps, and shapes unreliable — so treat a striking t-SNE picture as a question worth investigating, not an answer.

Skip to content