UMAP

UMAP (Uniform Manifold Approximation and Projection)

Definition

UMAP is a non-linear dimensionality reduction algorithm based on manifold learning and topological data analysis. It builds a fuzzy graph representation of the high-dimensional data and optimizes a low-dimensional layout to match it. Faster than t-SNE and parametric (can project new points).

How It Works

1. Build a fuzzy k-NN graph in high-dim space — each point connected to its k nearest neighbors with weighted edges based on distance
2. Construct a fuzzy simplicial set — a topological representation of the local structure
3. Optimize low-dim layout — find an embedding that minimizes cross-entropy between the high-dim and low-dim fuzzy graph structures

Properties

Property	Value
Type	Non-linear
Preserves	Local + some global structure
Speed	Fast (much faster than t-SNE)
Parametric	Yes — can project new points
Inverse transform	No (parametric UMAP extension exists)
Suitable for ML	Yes

UMAP vs t-SNE

	UMAP	t-SNE
Speed	Fast	Slow
New data	Yes (parametric)	No
Global structure	Partially preserved	Largely lost
Cluster spacing	More meaningful	Arbitrary
Hyperparameter sensitivity	Lower	Higher

Relevance to Search

UMAP’s parametric nature makes it more practical than t-SNE for embedding compression pipelines where new vectors must be projected at query time — t-SNE has no transform at all, so it cannot place a query in the indexed space. However, PCA remains the default for linear compression due to its speed, reversibility, and interpretability; UMAP’s cross-entropy objective still carries a local-structure bias, and its transform() is a fitted kNN-graph lookup rather than a matmul. See PCA vs t-SNE for Retrieval for the three-way verdict.

Key Parameters

• n_neighbors — controls local vs global structure balance (like perplexity in t-SNE)
• min_dist — minimum distance between points in low-dim space; lower = tighter clusters
• n_components — target dimensionality

Related Concepts

• Dimensionality Reduction — parent concept
• t-SNE — slower non-parametric alternative
• PCA — linear, even faster alternative
• Matryoshka Embeddings — training-time flexible-dimension approach

Articles

• PCA vs t-SNE vs UMAP - Visualizing the Invisible — comparison table with code