UMAP: Uniform Manifold Approximation and Projection
What you'll learn: How UMAP uses mathematical foundations from Riemannian geometry and fuzzy topology to preserve both local and global structure better than t-SNE.
The Core Idea
While t-SNE excels at preserving local neighborhoods (nearby points stay close), it often distorts global structure—distant clusters may appear at misleading distances. UMAP solves this by using a more sophisticated mathematical framework.
Think of it like this: Imagine mapping a curved surface (like Earth) onto paper. t-SNE is like drawing your neighborhood accurately but getting continent positions wrong. UMAP uses principles from Riemannian geometry (the math of curved spaces) to respect the manifold's true shape, keeping both your street and the continents positioned correctly.
Why UMAP Preserves Structure Better
Riemannian geometry lets UMAP model data as lying on a curved manifold, measuring distances along the surface rather than through space—like measuring driving distance instead of "as the crow flies."
Fuzzy topology handles uncertainty: instead of deciding "these points ARE neighbors," UMAP says "these points have a 0.8 probability of being connected." This fuzzy membership creates smoother, more stable representations.
Together, these foundations let UMAP:
- Preserve local structure like t-SNE (similar points cluster together)
- Maintain global relationships (relative distances between clusters are meaningful)
- Run faster on large datasets
- Scale better to more than 2-3 dimensions
Practical Advantage
Unlike t-SNE, where cluster distances are arbitrary, UMAP's embeddings preserve enough global structure that you can interpret the space between clusters. If cluster A is closer to B than C, that relationship likely reflects true similarity in the high-dimensional space.
Key Takeaway: UMAP uses Riemannian geometry and fuzzy topology to create dimensionality reductions that preserve both local neighborhoods and global structure, making it more mathematically principled and practically useful than t-SNE for many applications.