What is the difference between t-SNE and UMAP in practice?

t-SNE focuses on local structure, while UMAP preserves both local and global structure more efficiently.

t-SNE constructs probability distributions over pairwise distances and minimizes KL divergence, which strongly emphasizes local neighborhood preservation. However, it often distorts global structure and is computationally expensive. UMAP builds a fuzzy topological graph and optimizes a cross-entropy objective, preserving both local and some global structure while scaling better to large datasets.