How does UMAP differ from t-SNE for dimensionality reduction?

Question

AcadiFi · Accepted Answer

UMAP (Uniform Manifold Approximation and Projection) is a nonlinear dimensionality reduction method grounded in manifold learning and topological data analysis. Like t-SNE, it creates low-dimensional embeddings that preserve local structure, but it also does a better job preserving global structure, runs faster on large datasets, and is deterministic given a random seed. UMAP constructs a weighted graph of nearest neighbors in the original space using a fuzzy simplicial set representation, then optimizes a low-dimensional layout whose graph is topologically similar. The algorithm scales roughly linearly with sample size, making it suitable for hundreds of thousands of observations, whereas t-SNE slows quadratically. For a factor investing dataset at a firm like Beskarn Alpha with 80,000 stock-month observations across 25 factors, t-SNE might take hours and fit poorly; UMAP completes in minutes and produces embeddings where cluster distances are approximately meaningful. Analysts can feed UMAP outputs into downstream models (like clustering or classification) because the embedding preserves enough structure. Key parameters are n_neighbors (analogous to perplexity, controls local vs. global focus; try 10 to 50) and min_dist (controls how tightly points can pack). Smaller min_dist produces tighter clusters. ```mermaid graph LR H[40D data] --> U[UMAP
fuzzy graph] U --> L[2D layout] L --> C[Clusters] L --> M[Model input] ``` For CFA candidates, UMAP has largely replaced t-SNE in production quant pipelines because of speed, determinism, and global-structure preservation.

How does UMAP differ from t-SNE for dimensionality reduction?

Master Level II with our CFA Course

Related Questions

Practice Questions