Skip to main content
v2026.11,986 entries · CC-BY 4.0

Direct comparison

tSNE vs UMAP: Differences, Speed & Which to Use

tSNE vs UMAP: UMAP is faster at scale and may keep more global layout; t-SNE packs tight local clusters. Compare parameters, pitfalls and when to pick each.

Written and maintained by CASRAI Editorial Board

Last updated

How do UMAP and t-SNE compare side by side?

The table below compares UMAP and t-SNE across 17 dimensions, from full name to best fit.

Side-by-side comparison

DimensionUMAPt-SNE
Full nameUniform Manifold Approximation and Projectiont-distributed Stochastic Neighbor Embedding
IntroducedMcInnes, Healy & Melville, arXiv:1802.03426 (2018, rev. 2020)van der Maaten & Hinton, JMLR 9(86):2579-2605 (2008)
Core mechanismBuilds a fuzzy topological (simplicial-set) graph of nearest neighbors, then optimizes a low-dimensional layout to match itConverts pairwise similarities to probabilities in high-D and low-D (heavy-tailed t-distribution), minimizes KL divergence between the two
Local structureStrong -- visually competitive with t-SNE per the original paperStrong -- this is what t-SNE is specifically optimized for; produces tight, well-separated islands
Global structureAuthors' paper states it "arguably preserves more of the global structure" than t-SNEWeaker -- optimized for local neighborhoods; relative positions of distant clusters are not reliable
Between-cluster distancesNot a calibrated distance -- do not read gap size as a quantitative differenceDocumented limitation: cluster size and inter-cluster distance in the plot are not informative measurements
Key hyperparametern_neighbors (local neighborhood size) and min_dist (how tightly points pack)Perplexity (roughly, effective neighbor count), typically 5-50
Speed at scaleReported by its authors as consistently faster, with near-linear scaling and no penalty for embedding into more than 2-3 dimensionsSlower on large datasets; classic implementation scales less favorably, motivating Barnes-Hut and FIt-SNE approximations
Reproducibility across runsLayout can shift with random seed and hyperparameters; not guaranteed identical run-to-runSame caveat -- known to be sensitive to random initialization and perplexity choice
Feeding a downstream algorithmIts neighbor graph is commonly reused directly by downstream tools (e.g., graph-based clustering)Typically used only as a terminal visualization, not as input to further steps
Typical default in single-cell pipelines (2026)Most common default visualization step in current single-cell toolkitsStill widely used, especially where tight local separation is the priority
Axis interpretabilityUnitless, non-interpretable individually -- no loading/variance-explained meaning like a PCA biplotSame -- axes carry no independent quantitative meaning
Neighborhood-size defaultn_neighbors = 15 in umap-learn; low values emphasize very local structure, high values capture broader structure (umap-learn documentation)Perplexity has no universal best value; the 5-50 range is only a starting point and it must stay below the number of points (Wattenberg et al., Distill 2016)
Cluster packing controlmin_dist (default 0.1) sets how tightly points may pack; low values give clumpier embeddingsNo equivalent; tightness follows local density and perplexity
Cluster size in the plotNot a count or variance measure; do not compare cluster areasDense clusters are expanded and sparse ones contracted, so relative size is not readable (Distill 2016)
Convergence checkRe-run with another seed and n_neighbors before trusting an arrangementStopping before convergence can show pinched or point-like shapes that look like real structure; check that the layout has stopped changing
Best fitLarge datasets and when the embedding or its graph feeds later stepsSmall to moderate datasets where crisp local separation is the goal

Ask CASRAI · free to try

Ask about tSNE vs UMAP: Differences, Speed & Which to Use

2 free questions a day, no account.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

Common questions

Common questions about UMAP vs t-SNE

Does UMAP preserve global structure better than t-SNE?

+

That is the claim UMAP's original authors make in the introducing paper -- that it "arguably preserves more of the global structure" while remaining visually competitive locally. Treat it as the algorithm's design goal and the authors' reported result, not an independently settled distance-accuracy guarantee.

Can I trust the distance between two clusters in a UMAP or t-SNE plot?

+

No, not as a quantitative measurement. Both algorithms optimize local neighborhood preservation; the gap between clusters and the apparent size of a cluster are artifacts of the optimization and hyperparameters, not calibrated distances or sample counts. Verify any global pattern against the original high-dimensional data before reporting it as a finding.

Is UMAP always faster than t-SNE?

+

It is reported as consistently faster in the original comparison, and that holds up in common practice, particularly as dataset size grows into the tens of thousands of points and beyond. On small datasets the difference is usually negligible.

Should I run PCA before UMAP or t-SNE?

+

Yes -- this is standard practice, especially for high-dimensional data like single-cell expression matrices. Reducing to a moderate number of principal components first cuts noise and computation before either nonlinear method runs its neighbor search.

Do perplexity (t-SNE) and n_neighbors (UMAP) mean the same thing?

+

They play an analogous role -- both roughly set the effective local neighborhood size -- but they are not mathematically identical, and default values differ by implementation. Changing either meaningfully changes how tight or diffuse the resulting embedding looks.

What is the difference between UMAP and t-SNE?

+

Both turn high-dimensional data into a 2D or 3D map by preserving local neighborhoods. t-SNE minimizes the KL divergence between high- and low-dimensional similarity distributions and is tuned by perplexity. UMAP builds a fuzzy nearest-neighbor graph and is tuned by n_neighbors and min_dist. UMAP is usually faster at scale and its authors report it keeps more global structure; t-SNE often shows tighter local clusters.

Which is better, t-SNE or UMAP?

+

Neither is better in every case. UMAP is the usual default for large datasets and when the neighbor graph feeds later steps; t-SNE is a good choice for small or moderate datasets where tight local separation matters. For either, check that the structure survives a different seed and different neighborhood settings.

Is t-SNE or UMAP better for single-cell RNA-seq?

+

UMAP is the more common default visualization in current single-cell toolkits, mainly for speed and its reuse of the neighbor graph. t-SNE is still used when local cluster separation is the priority. Cluster calls should come from a clustering method such as Leiden clustering, not from how the plot looks.

How do I choose perplexity for t-SNE?

+

There is no single best value. The 5-50 range is a common starting point, perplexity must be smaller than the number of points, and low values can create random-looking clumps while high values can merge clusters. Compare several values and keep only patterns that persist.

What do n_neighbors and min_dist do in UMAP?

+

n_neighbors (default 15 in umap-learn) controls the balance between local and global structure: low values emphasize fine local detail, high values broader structure. min_dist (default 0.1) controls how tightly points are packed; low values give clumpier embeddings.

UMAP vs PCA vs t-SNE: when should I use each?

+

PCA is linear, fast and its axes have variance-explained meaning, so it suits denoising and a first look. t-SNE and UMAP are nonlinear and built to reveal local neighborhood structure in a plot, with uninterpretable axes. A common workflow runs PCA first, then UMAP or t-SNE on the leading components.

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →